Last week, I was at the University of Chicago talking about all things AI. After the panel, I struck up a conversation with an audience member who worried that frontier AI companies were racing to the bottom when it comes to AI safety. What he meant is that the competition to release more capable models might create pressure to skimp on testing.

I think he was right but not for the reasons he imagines. AI has been involved in a race to the bottom, but it’s been towards more alignment. The competitive pressure to release new models has also created powerful incentives to build better alignment tools.

Indeed, AI is unique in that it is spawning a new class of models to make the tech better. The Petri auditor agent, for example, subjects a target model to realistic, multi-turn scenarios and tries to elicit deception, misuse, sycophancy, sabotage, and a range of other misaligned behaviors. A separate judge model then scores the interaction. Jan Leike, lead of alignment at Anthropic, explained the power of these auditor models: “for the first time we have an alignment metric to hill-climb on. It’s not perfect, but it’s proven extremely useful for our internal alignment mitigations work.”

Results from the Petri auditor agent showed a striking trend over 2025. Models got substantially more aligned as misalignment dropped. Anthropic wasn’t the only AI company doing better. Google and OpenAI have also made significant progress. xAI is a different story altogether.

Alignment is hard and incomplete, to be sure, but the major labs are making progress. As Jan Leike explained in a post titled “Alignment is not solved: But it increasingly looks solvable,” the scale-up is “by no means finished, but at this point we have some solid evidence that we can manage these misalignments while scaling up.”

I’ve written before about Aaron Wildavsky’s core argument in Searching for Safety, that safety is “largely an unknown for which society has to search.” A policy or intervention might make us safer, or it might not. Often, he wrote, we only learn which is which after the fact. The same is true of AI. Safety is sought by searching for it, by building systems, probing them, auditing them, red-teaming them, and learning.

Wildavsky drew a distinction between strategies of resilience and anticipation. Anticipatory strategies try to predict and prevent harm before it occurs while resilient strategies build the capacity to absorb shocks. Wildavsky’s bet was on resilience because it compounds. Each correction adds to a stock of knowledge that makes the next correction easier and faster.

Aaron Wildavsky’s insights on risk seem to be holding.

The open question is whether that loop can keep pace with capability gains, especially with tech like Anthropic’s Mythos. Misalignment scores fell sharply in 2025 even as model capabilities rose, which is an encouraging trend. But the relationship between capability and misalignment risk is not linear and there’s no guarantee the trend will hold at the next order of magnitude of capability. Still, Wildavsky’s insights are powerful as they encourage us to continue the search and build better detection. He would also say that stopping the search, on the grounds that the next step might be dangerous, forfeits the knowledge needed to manage that danger.

If the future is radically uncertain, then the rational response is not to obsess over unconditional doom forecasts. It is to look for variables we can actually move. Automated auditing is valuable precisely because it turns alignment into something like an engineering target. The real race, then, is not to abandon the search for safety, but to learn fast enough to keep alignment ahead of capability.

Until next time,

🚀 Will

At AEIdeas, I recently wrote about “The Political Backlash to Data Centers,” where I pulled together all of the polling on data centers and the state bills that would impose data center moratoria. In mid February, I submitted testimony for New Hampshire’s HB 1589, which would impose social media interoperability. It’s a topic I’ve written about a lot over the years and this testimony summarizes that research.

I’ve also been thinking about “The Coming Fight to Define the Agentic Web.” As AI agents start to act and shop on our behalf, major platforms are deciding whether to treat those agents as customers or intruders. This fight will shape commerce infrastructure for years to come.

Finally, in late January, I wrote “Jagged Intelligence, Jagged Adoption,” riffing on AI pioneer Andrej Karpathy’s notion of jagged intelligence to argue that business level adoption of AI will be jagged as well. To better understand this jaggedness, I put together a table that compiled major enterprise surveys on AI adoption and ROI. Across studies, only 5-15% of organizations report significant, measurable ROI from AI initiatives.

On the AEI panel last week titled “Moral Questions in the Age of AI,” Duke Law Professor Nita Farahany relayed the moment she realized large language models (LLMs) can produce outputs that rival a human. It was a moment people are increasingly experiencing, where the categories we use to understand ourselves are destabilized. When a machine produces an essay indistinguishable from one written by a thoughtful person, we are first taken aback by its capabilities. But then we are left speechless that we cannot articulate what the difference between us and them is.

She got AGI-pilled.

While the panel covered a lot of ground, it was a missed opportunity to discuss bigger ontological, metaphysical, and epistemological questions that AEI can uniquely cover. The arrival of systems that can write and reason forces a much needed confrontation with our settled categories. LLMs have deepened our understanding of the way things are, the nature of communication, what is distinctively human, and what separates mind from machine. But by not tackling these issues head on, the panelists ended up dancing around fundamental philosophical questions.

What does it mean that matrix multiplication employed in transformers, trained on neural nets, boosted by reinforcement learning, and all the rest can create something that closely approximates human-level writing and reasoning? And if a machine can do the things we considered distinctively human, what is distinctively human?

Gideon Lewis-Kraus’ widely shared New Yorker piece, “What Is Claude? Anthropic Doesn’t Know, Either,” is about one stream of this question, which can be understood tracking the mechanistic interpretability work at Anthropic. Lewis-Kraus writes:

Language is, or rather was, our special thing. It separated us from the beasts. We weren’t prepared for the arrival of talking machines. Ellie Pavlick, a computer scientist at Brown, has drawn up a taxonomy of our most common responses. There are the “fanboys,” who man the hype wires. They believe that large language models are intelligent, maybe even conscious, and prophesy that, before long, they will become superintelligent. The venture capitalist Marc Andreessen has described A.I. as “our alchemy, our Philosopher’s Stone—we are literally making sand think.” The fanboys’ deflationary counterparts are the “curmudgeons,” who claim that there’s no there there, and that only a blockhead would mistake a parlor trick for the soul of the new machine. In the recent book “The AI Con,” the linguist Emily Bender and the sociologist Alex Hanna belittle L.L.M.s as “mathy maths,” “stochastic parrots,” and “a racist pile of linear algebra.”

And then he later says:

We don’t know if it makes sense to call them intelligent, or if it will ever make sense to call them conscious. But she’s also making a more profound point. The existence of talking machines—entities that can do many of the things that only we have ever been able to do—throws a lot of other things into question. We refer to our own minds as if they weren’t also black boxes. We use the word “intelligence” as if we have a clear idea of what it means. It turns out that we don’t know that, either.

Those early lines got me: Language is, or rather was, our special thing. It separated us from the beasts. For decades, evidence from animal cognition has chipped away at the idea that language cleanly separates us from the rest of the living world. Sperm whales are the strongest case.

Recently, 19th Century whaler logbooks from the North Pacific, one of the last whale breeding grounds to be hunted, were digitized and researchers found that the ships’ strike rate fell by 58% in less than two and half years after the whalers first arrived in the region. The dramatic decline suggests that there was some kind of information sharing among whale groups. Whalers have long talked about codas but it wasn’t until the 1950s that science finally recognized sperm whales communicate through click patterns that we now know vary by clan and can be carried over great distances through the SOFAR channel.

The whale evidence matters because it collapses the clean version of the language argument. If cultural transmission, combinatorial structure, and long-distance communication exist outside human language, then language was never the bedrock we assigned it. The question was never whether we could talk. It was whether talking, by itself, explained anything about mind, meaning, or moral standing. We just avoided the uncomfortable conclusion because no other species was making us face it directly.

But machines have accelerated that erosion. They do not merely communicate, they are able to participate in discourse. They are drafting contracts, writing poetry, summarizing case law, and simulating empathy. We defined intelligence operationally for decades, then built systems aimed toward those operations and were shocked that they worked. The shock is that our definitions were always more mechanistic than we admitted.

Indeed, is there any reason that the most useful and widely adopted form of AI is the chatbot?

Dialogue is the foundation of philosophical thinking. Famously, it was Protagoras’ dialogical way of arguing that sparked Socrates to search for a firm footing of truth in the Socratic dialogues. The dialogic form turns out to be generative for machines for the same reason Protagoras thought it was generative for humans. Thinking, at its most productive, is an argument with itself.

But if gradients can generate essays that pass for thoughtful reflection, then perhaps some portion of what we call reasoning is statistical pattern completion at scale. That does not mean human thought reduces to linear algebra. It does mean that at least part of what we experience as insight may emerge from processes that are less mysterious than we suppose.

Some have retreated to the safety of taste. But in a must-read essay, Will Manidis eviscerates that refuge: “No one proposed it. No one even had to. It was a clean and easy answer to the question everyone in technology has been asking: what are humans for once the models get good enough?”

Manidis is right in the broad strokes: Taste as we know it arrived when the patron left the room. In the Middle Ages into early Modernity, creative production was a negotiation between patron and maker, oriented toward something transcendent. The Sistine Chapel happened because Julius II and Michelangelo fought. Work emerged from friction, not selection. But when modern production markets arose in the early 20th century, capital and labor got unbundled. The collector replaced the patron. Now art is evaluated by the finished work or the process. Tom Wolfe documented this in The Painted Word, while Pierre Bourdieu produced the academic tome.

Manidis is right that AI again flips this relationship. We should see AI as giving us the ability to make something transcendent. But scarcity was never just an economic necessity, it was also a condition of production. A book took years to write because the author had to develop something worth saying. The difficulty was not incidental to the value. It was constitutive of it. What happens to discourse when the friction of writing and artistry is gone?

Matthew Crawford, author of Shop Class As Soulcraft, was on the AEI panel as well. He should have been asked how AI alters the thesis of his book. The mechanic’s knowledge of an engine is not a set of propositions about engines. It is a practiced relationship with a kind of object, accumulated through friction and failure and learning. A model trained on every repair manual might be able to perfectly tighten a nut but it still cannot express the feeling when a bolt is about to strip.

Indeed, exploring this topic would help partly answer the question Mike Bird of The Economist asked,

What’s the best understanding of why AI-generated writing is still so janky (e.g. better than many students, or non-professional writers/academics, but usually not nearly as good as professional writers)?

AI produced writing is good but it still lacks the voice of an author or the vision of an auteur. You can try to prompt it to create output in a certain style like Joan Didion or Tom Wolfe but it still lacks their cadence. And it lacks that cadence because the gallop of writing is a record of how a particular mind moves.

These machines have also reopened old questions: What does it mean that different AI architectures, trained on different data for different tasks, are independently arriving at similar internal representations of the world?

Vision models and language models, built differently and trained differently, appear to be converging on the same underlying structure of concepts like shape, color, and spatial relations. It’s called the Platonic Representation Hypothesis. Models were never told to build these structures. They emerged independently from exposure to data. If true, it could mean models are latching onto something real about the structure of reality. If true, it touches almost every major fault line in modern philosophy from debates over rationalism to empiricism to structuralism.

The rationalist reads this as evidence that certain structures are prior to experience, that the mind latches onto form because form is really there. The empiricist has a harder time. If two systems trained on entirely different sensory inputs converge on the same representations, is it still true that structure is purely a product of experience? The structuralist, meanwhile, finds it vindicating. Reality has an architecture, and any sufficiently powerful learning system will find it. None of these positions is formally settled but the evidence is now empirical, which fundamentally changes the terms of the debate.

Indeed, if thinking is more mechanistic than we assumed, one question becomes clearer: What is it that cannot be captured by representation and prediction alone?

For a panel with at least one avowed Catholic, I was surprised that there was no mention that LLMs might point to a seemingly radical position in this materialist world, the existence of a spirit, of a soul. LLMs’ ability to reason suggests that what is distinctively human has to lie elsewhere. The soul, in the classical Christian account, was never just about the ability to produce language. It was long understood to be the seat of will, intentionality, moral responsibility, and the capacity for the good, endowed by the Creator. If machines can simulate thought, perhaps what distinguishes us is this first-person perspective ordered toward truth and the good. Humans aren’t systems that model the world. They are individuals that stand in moral relation to it.

The panel was also a chance to probe that uncomfortable trend: Why does everyone bring up the Terminator scenario? What is it about the nature of humans that our first creation is going to kill us?

If you spend any time in the AI debates, you will encounter people who are convinced that AI will kill us all. Our history is littered with such worries, about nuclear weapons, industrial pollution, and engineered pathogens. But there is something deeper about the apocalyptic tone with AI. We do not merely worry that AI will malfunction. We worry that it will supersede us, make us obsolete, or turn us into paperclips. We project onto it the very tendencies we know lurk in ourselves. When humans encounter something weaker, we exploit it. When we encounter something stronger, we fear domination. The AI discourse reads at times like a mirror held up to ourselves, a retelling of the Tower of Babel in Genesis 11, or Prometheus, or Frankenstein. The fear of AI killing us is a dramatized version of a more basic truth that we have yet to grasp, that we are creatures capable of building beyond our wisdom. The Fall of Man in Genesis 3 has been reborn for the AI age.

While my day to day work is in AI policy, the AI debate is mostly conducted at the wrong altitude. We argue about capabilities, safety benchmarks, and economic disruption, which are all important, and yet still, the deeper questions go unasked. What are we if AI can approximate reason? What is thought if gradients can reproduce it? What is left of human distinctiveness once we strip away every faculty a machine can replicate? These are not rhetorical provocations. They are the questions that the AEI panel should have been asking. These are the questions we should be asking.

Until next time,

🚀 Will

Last month, Senators Elizabeth Warren, Chris Van Hollen, and Richard Blumenthal sent an angry letter to the CEOs of leading tech companies, singling out data centers for climbing energy bills. They are big and energy-hungry, after all, and it seems that tech companies are “passing on the costs of building and operating their data centers to ordinary Americans as AI data centers’ energy usage has caused residential electricity bills to skyrocket in nearby communities.” The numbers sound terrifying:

As a result of the combined energy needs of AI data centers and cryptocurrency miners, electricity bills are estimated to rise 8% averaged nationwide by 2030 and up to 25% in states like Virginia with a high concentration of data centers.

Political pressure has only intensified since then. Just this week, President Trump called on tech companies to “pay their own way” and not burden local ratepayers with infrastructure costs. Microsoft quickly pledged not to seek property tax breaks or allow its facilities to drive up electricity rates. This high-level attention reflects genuine concern about surging electricity costs but it also risks cementing a misleading narrative.

While a 25 percent price hike sounds dire, when you look at actual estimates from Virginia, this increase represents the top-end estimate, with a middle range projection of a 7 percent increase. It’s not nothing, but it is not evidence of a price spiral from AI. The real kicker is that this same research found that data centers have paid their way, though this might not continue if policy doesn’t change.

Empirical analysis on this question is clear. Data centers haven’t been the big driver for rising energy prices. It’s been inflation. Counterintuitively, this research finds that “the greatest price increases typically exhibited shrinking customer loads.” Data centers use a lot of power, but whether that power consumption automatically translates into higher bills for everyone else depends on regional policy choices and how the individual contracts are structured.

Energy cost projections are notoriously difficult to get right, and advocates frequently exploit this uncertainty by embedding their preferred policy solutions within seemingly technical assumptions. To understand just what was being predicted, I followed the 25 percent increase stat in the letter to a New York Times article and then to a series of Carnegie Mellon University (CMU) reports. The numbers are detailed here:

Home to the nation’s largest data center concentration, the region is projected to require an additional 100 TWh of electricity by 2030. Our analysis shows that, under current policies, more than 25 GW of aging and costly coal plants would continue operating largely to meet the added electricity demands from data centers, plants that would otherwise be on track to retire. This added generation drives a projected cost increase of more than 25% in Central and Northern Virginia—the highest regional increase in the model. Under these conditions, price spikes similar to those seen in the December 2024 PJM capacity market auction may become more common.

Still, it doesn’t make much sense to me why the CMU report assumes that “increasing generation from older coal plants—while expensive to operate—is still less expensive in the short-run than building new gas.” Why would expensive coal generation be the marginal resource rather than new natural gas combined cycle plants? I asked Claude about these assumptions and it responded, “You’re right to be skeptical. This reads more like advocacy research designed to support predetermined policy conclusions than a neutral capacity expansion analysis.”

Virginia is home to about 13 percent of all global data centers, with numbers expected to grow substantially in the near term. To plan for this future growth, Virginia’s Joint Legislative Audit and Review Commission (VJLARC) commissioned a report from E3, a widely respected energy consulting firm, to figure out what data center growth would mean for the state. In December 2024, the report was released and it was accompanied by a review of electric infrastructure and rate impacts from projected data center growth.

While these forecasts are never perfect, they estimated load growth demands assuming no new data centers, unconstrained data center growth, and a moderate projection. Then these load growth scenarios were varied assuming that the state’s utilities did or did not comply with the Virginia Clean Economy Act’s (VCEA) requirement of 100 percent zero carbon generation portfolios by 2050. Assuming that Virginia doesn’t meet VCEA standards, which the law allows for, and assuming unconstrained data center growth, “Virginia is projected to add significant amounts of new gas capacity at an accelerated rate in the near term, compared to the no growth case.” As the graph below illustrates, coal is nowhere near the level that the CMU report assumes. Though, to be fair, this may come from net energy imports from nearby regions.

In spite of these differences, the VJLARC report estimated that 2030 energy prices could see a rise of 25 percent, similar to the CMU report. This was a top-end estimate from the unconstrained scenario, assuming that data centers could be “sited, built, and interconnected as fast as the market desires.” But, as the report continued, “in practice, constraints on the pace of infrastructure development may limit how quickly these facilities can add electric demand to the system.” Instead, if there were constraints in data center growth, prices would be a much more manageable 7 percent. Energy policy is hardly immune to politics. Advocates lead with the most alarming figure while burying the assumptions that drive the numbers. The headline number is technically defensible, but only under an extreme scenario.

But the real narrative violation is that in reviewing the past few years of price increases, the VJLARC report found that “current rates appropriately allocate costs to the classes and customers responsible for incurring them, including data center customers.” That finding directly undercuts the popular claim that data centers have been quietly shifting their costs onto ordinary ratepayers.

These results are corroborated by a new paper from researchers at the Lawrence Berkeley National Laboratory (LBNL) looking at price increases from 2019 to 2024. They reached a conclusion that cuts against the current thinking about data centers: load growth is associated with lower retail prices. To be fair, the LBNL paper and the ratepayer review both look backward to understand what drove retail electricity prices, so these findings might not hold in the future.

This slide deck offers a great overview of all the work behind the LBNL paper, though there are six trends worth highlighting because they explain a lot of what has happened with price hikes:

1. National-average nominal retail prices have largely tracked economy-wide inflation in recent years, so real prices have been flat;

2. Residential customers have faced larger recent price increases than have commercial and industrial customers;

3. State-level retail electricity price trends vary, with some states’ prices rising much faster than inflation;

4. Natural disasters, extreme weather, and wildfire mitigation have significantly increased prices in some states;

5. Natural-gas price fluctuations have been among the largest factors impacting year-to-year variations in retail electricity prices; and

6. Many state-level renewable portfolio standards—which require that energy is sourced from renewable energy sources like wind, solar, and geothermal—have increased retail electricity prices.

I think most people would be surprised to learn that states with shrinking load tended to see higher prices. This helps to explain, partly, why California is having an explosion of energy prices. Rooftop solar mandates are driving down use, which means that fixed grid costs were spread over fewer kilowatt-hours. When new load arrives and can be served largely with existing infrastructure, it dilutes embedded fixed costs, pushing average prices down. It’s this chain of events, the LBNL paper details, that helps to explain why a 10 percent increase in statewide load is linked to a 0.6¢/kWh reduction in prices.

But what might we expect in the future as data centers continue to expand? Amazon commissioned E3 to study this, and their framework helps to clarify what might happen in a region in one of three different scenarios:

1. Small increases that the grid can handle — This describes most of the recent data center growth. When electricity demand rises, utilities can serve new customers using power plants and transmission lines that already exist. And if the incentives are right, everyone’s average price can fall because fixed costs get spread across more customers.

2. Growth that requires new infrastructure — Sometimes demand grows enough that utilities need to build new power plants or transmission lines. In this case, what happens to residential electric bills depends entirely on who pays for this new infrastructure. If the new customer, like a data center, pays the marginal infrastructure costs, the effect on other ratepayers is neutral. If the new customer also helps cover embedded fixed costs, average prices for residents decline. If those costs are instead socialized, prices for other customers rise. It depends entirely on the cost allocation equations.

3. Massive growth that transforms the whole system — Occasionally, electricity demand expands so much that it rivals the size of the existing grid itself. Meeting this demand requires utilities to build new infrastructure at an unprecedented scale. In these transformational load scenarios, the traditional approach to setting electric rates breaks down and avoiding unfair cost-shifting requires custom contracts that spell out exactly who bears which risks.

What I take away from the totality of this evidence is this: whether ordinary consumers bear the costs of data center growth is a policy choice. Data centers do not inherently raise residential electricity bills. They do so only under rate designs that allocate infrastructure and system costs in ways that allow residential customers to subsidize large, high-load users. In Georgia, for example, the state’s largest utility is set to double their total generation over the next five years while also cutting the average household electric bill by more than 50 percent. And rightly, the VJLARC report suggests “establishing a separate data center customer class, changing cost allocations, and adjusting utility rates more frequently could help insulate non-data center customers from statewide cost increases.” Read this way, the projected price increases might be better understood as a warning about cost allocation methods, which are opaque and highly technical as this 79-page presentation underscores.

Energy contracts also have a part to play. Many contracts include capacity reservation fees. Data centers pay for reserved power whether or not they use it, helping utilities plan infrastructure investments without passing emergency capacity costs to other ratepayers. Some agreements include interruptible service clauses. The data center gets discounted rates but must curtail usage during peak demand periods. Economic development agreements are also common where data centers will fund energy infrastructure upgrades. These arrangements can create win-win outcomes.

The senators who wrote to Amazon aren’t entirely wrong to be concerned. Virginia’s electricity demand is genuinely surging, and if that growth is managed poorly, residential customers could get stuck with the bill. And data centers make for a convenient target. They’re big, they’re owned by wealthy tech companies, and AI is everywhere in the news. But if we’re going to fix electricity pricing problems, we need to understand what’s actually causing them. Blaming data centers for rising electricity bills is easier than reforming how we allocate infrastructure costs, but only one of those approaches will actually help ratepayers.

Until next time,

🚀 Will

The project came about because I heard that one statement so often: We put a man on the moon, why can’t we do this as well? The Apollo program maintains a mythical status in tech policy, not only as proof that the government can do big things, but also as evidence that scientific ambition plus money equals results. Be it climate change or cancer or AI safety or semiconductor manufacturing, it is imagined that any sufficiently urgent problem can be solved if we just spend enough money and resources to moonshot it.

I went into this project just trying to understand what made Apollo successful. What came out of this was a paper that punctures the myths without diminishing the achievement. Apollo was a triumph, to be sure, but not for the reasons people usually imagine. Apollo’s success was largely due to management and oversight.

For one, I didn’t truly understand the historical context of Apollo until I started laying out all of the key dates. A little over a week separated Yuri Gagarin’s first spaceflight in April 1961 and the Bay of Pigs fiasco. These two events, a Soviet triumph and an American embarrassment, pushed President John F. Kennedy to back an ambitious space mission. Not long after, a moon landing emerged as the preferred option and by the end of May, Congress had approved the idea.

This compressed timeline got me thinking about the conditions that cause government projects to move quickly, which I take up in the final sections of the paper. It’s there that I compare Apollo to two recent programs: Operation Warp Speed (OWS), which accelerated the development of COVID vaccines, and the Broadband Equity, Access, and Deployment (BEAD) program, which funds nationwide broadband deployment. BEAD has been a boondoggle while OWS is seen as a success, and some of that is due to the speed. OWS was sketched out in mid-April 2020 and funded by May. It was meant to move quickly to address a crisis, much like Apollo. I know there’s interest in reinvigorating industrial policy, but at least for these two projects, external circumstances forced action, which is hard to replicate without a crisis.

Another major misconception about Apollo is that it was a science project. By 1961, all the pieces were in place to plan a mission to the moon. Missile programs had proven orbital mechanics, rocket propulsion, and materials science. Saturn’s massive engines were about to be delivered to NASA, following plans for a moon mission that were laid out back in 1957. There were tough engineering problems to solve, but as I wrote, all the work in the years leading up to 1961 “allowed NASA to set concrete milestones, predict resource requirements, and maintain accountability in ways that pure research projects cannot. The distinction explains why Apollo could promise a specific timeline to land on the moon before the decade was out, while true scientific breakthroughs were and remain fundamentally unpredictable.” Modern discussions about moonshots often misapply this model to projects requiring fundamental scientific breakthroughs.

Probably the most surprising thing I learned is that Apollo came exceptionally close to missing the decade deadline. By 1963, Apollo was in trouble and schedules were slipping, so much so that estimates put the earliest moon landing in 1971. So George Mueller (pronounced Miller) was brought in to manage it all. More than anyone else, Mueller made Apollo successful.

As I wrote:

Mueller was a systems engineer who had honed his management skills in the Air Force’s missile program. Among other changes he made when he took control, Mueller implemented a program office organization within NASA’s existing institutional structure that remained throughout the Apollo program. Before Mueller, NASA’s organizational structure was functionally divided, with separate engineering divisions, manufacturing groups, and testing departments that made coordination across a complex program difficult. Mueller introduced a modern management approach that gave program offices direct authority over all aspects of their missions, cutting across the traditional functional divisions. He also streamlined operations in Project Gemini, which was established to develop spaceflight capabilities to support Apollo, and increased the pace of launches so that lessons could be learned and then applied to Apollo.

Most importantly, Mueller pushed the all-up testing approach for the Saturn V; instead of incrementally testing one stage at a time, he insisted on flying all stages together. To veteran rocketeers like Wernher von Braun, this sounded reckless. In practice, however, it was the only way to compress the timeline enough to meet Kennedy’s deadline. In effect, Mueller parallelized the problem, which is exactly what happened in OWS, as I explained:

The government backed multiple approaches, including mRNA and typical platforms, to maximize the chances that at least one would succeed. Some companies, such as Moderna and Novavax, received direct federal funding for research, clinical trials, and manufacturing scale-up, while Pfizer-BioNTech entered advance purchase agreements. In parallel, OWS poured funds into expanding manufacturing capacity by helping build production lines and secure raw materials before vaccines were approved. The US Army Corps of Engineers, for example, helped expand factory capacity. Effectively, the government paid for facilities and production expansion for vaccine candidates it might never select.

For policy wonks, the single most interesting book I read for this project was Arthur L. Slotkin’s “Doing the Impossible: George E. Mueller and the Management of NASA’s Human Spaceflight Program.” Near the beginning, Slotkin details Mueller’s overall management philosophy, which Mueller explained in a series of talks given in the late 1960s:

The criteria necessary to manage large scale R&D programs required the objective to be within the state of the art without having to reach for performance, and be achievable in a reasonable timescale. Activities required to be clearly defined for each organizational element, with engineering and design aiming for high reliability, and manufacturing providing adequate inspection and quality control. But he cautioned that reliability has to be designed into an object; it “cannot be built into a badly engineered design,” it can only be achieved by having a sound design to begin with, followed by careful manufacturing and testing. Since change is part of R&D, the design must tolerate changes. On the other hand, he warned, “we must guard against change leading to overdesign.”

The uncomfortable implication of all this is that Apollo doesn’t support many of the arguments it’s routinely used to make. Money cannot solve coordination problems. Scientific uncertainty cannot be scheduled away by political will. The government cannot replace private industry. What it shows is that we already know how to run complex technological programs but only if policymakers accept a set of unglamorous but essential requirements:

• Ensure that the program’s objective is technologically possible without resorting to exceptional performance;

• Rely on the ingenuity of private business;

• Move quickly from proposal to contractor selection;

• Convert sequential problems into parallel problems when possible;

• Adopt management techniques that have been successful in similar domains; and

• Implement strict planning and oversight systems.

Apollo still has a lot to teach us, and it’s not that we need to dream bigger. Rather, it’s that we need to focus more attention on better management. The full paper is here, so let me know what you think!

Until next time,

🚀 Will

Fifty years ago today, the SS Edmund Fitzgerald sank. For any child of the Great Lakes growing up in the 1990s, as I did, the event is infamous; not only because Gordon Lightfoot’s song was constantly played on soft rock stations. The sinking of the ship was a marker between the Midwest as a Factory Belt and the Midwest as a Rust Belt. Hauling ore from the Iron Range in Minnesota to iron works in Michigan, those famous last lines from Captain McSorley captured perfectly the Midwestern reserve in the face of adversity: “We are holding our own.”

Having spent a lifetime reading about the shipwreck, I now see the event as a true tragedy. Only a few miles behind the Edmund Fitzgerald was the SS Arthur Anderson, which survived the storm that produced wind gusts over 80 miles per hour and waves of up to 25 feet. The storm was a feared November Witch, created when a low pressure zone over the region pulls in cold from the Arctic and warmth from the Gulf, resulting in hurricane-like conditions on Lake Superior. But neither one of those ships should have been in the path of the storm. Had weather forecasting been better, the tragedy never had to occur.

The National Weather Service’s (NWS) long range forecast on November 10, 1975 predicted that the storm would pass at the southern portion of the Lake. With this in mind, the Arthur M. Anderson and Edmund Fitzgerald took the trip across the lake following the normal Lake Carriers Association route. That course, however, placed them directly in the path of the storm. Meanwhile, Captain Paquette of the SS Wilfred Sykes had been tracking the system for a couple days and concluded that it would pass over the eastern part of the Lake. Paquette was known for keeping meticulous notes and tracking weather. So, instead of making the trek, he took refuge in Thunder Bay, Ontario, during the worst of the storm, safely out of harm’s way.

In the early afternoon, the storm picked up and both the Anderson and the Fitz made way to the Canadian shoreline, which would have meant safety had the low pressure zone been passing to the south. Instead, it exposed them to the absolute worst of the storm. By the late afternoon, the storm dramatically intensified to hurricane-like conditions.

Joe Warren was a 23-year-old deckhand on the Anderson and he later recounted in So Cold A Sky: Upper Michigan Weather Stories what he experienced: “I had never seen the flex and the bend or heard steel groan in such a horrific way. The waves had turned into mountains.” His ship slowed down while the Edmund Fitzgerald continued chugging ahead, trying to reach the safety of Whitefish Bay.

As this was happening, the chief engineer of the Anderson joined Warren on deck. “I will never, never forget what he said,” recalled Warren. “He said, ‘That old man [referring to Captain McSorley of the Fitzgerald] is either going to put ‘er on the bottom or he’s going to tear the engine out of ‘er.’ At least Cooper, my captain, checked down.”

Modern analysis of the storm found that the Fitz sailed directly into the worst part of the storm when it was at its most intense. The Anderson last heard from the Fitzgerald at 7:10 PM, which is when the storm would have been at its peak. Tom Hultquist, who conducted the analysis at NWS, explained his findings in So Cold A Sky, saying, “It ended up in precisely the wrong place at the absolute worst time.”

But given the advancements in technology, it’s highly unlikely that such a tragic loss of life would happen today.

In 1975, weather forecasting was in a dark age. At the time, there weren’t even wave detecting buoys on Lake Superior. Forecasts were updated daily and meteorologists depended entirely on voluntary ship reports to track conditions over the lakes. Ships received weather communiques through VHF radio broadcasts, which crew members would have to manually record and plot. Missing a single broadcast could mean missing critical forecast updates.

Ironically, 1975 was the year that this would begin to change. On October 16, 1975 the first Geostationary Operational Environmental Satellite (GOES-1) was launched into space, returning its first images late in the month. GOES-1 was the first in a series of dedicated geostationary satellites for real-time weather tracking that have revolutionized the field. Today, GOES-18 and GOES-19 actively monitor the globe for changes in weather. Steady advances in technology allow for a range of operations today with GOES that were unthinkable even under GOES-1, including fire detection, air quality warnings, fine grained detection of fog, and enhanced tornado warning lead time, among others.

In the mid 1970s, the NWS’s models were simple and unreliable. Lake Superior, in spite of being the size of South Carolina, had just 3 grid points, or data points, for weather models. By the early 2000s, this had increased to 420 grid points. Current models are even more sophisticated, having moved on from single layer models to 127 vertical levels and sophisticated parameterizations. Forecasters in 1975 struggled to predict conditions 24 hours ahead with confidence, which meant that ships would often sail in horrible conditions. Modern systems reliably forecast conditions five to seven days out.

While weather forecasting was primitive in 1975, wave forecasting was nonexistent. Lake Superior is the second largest lake in the world by surface area and is the largest of the five Great Lakes. With depths reaching 1,333 feet, Lake Superior behaves more like a sea than a lake. So, it took a revolution built on data to understand and predict this body of water. In 1974, the Great Lakes Environmental Research Laboratory (GLERL) was established and by the early 1980s, the lab had developed technology for producing wave forecasts for the Great Lakes. Current wave forecasts are driven by WaveWatch III, a third generation wave prediction model. This model produces twice-daily forecasts and is part of a suite of models that have been developed to predict conditions on the lakes known as the Great Lakes Coastal Forecasting System (GLCFS). By integrating lake currents, temperature, wind, waves, and ice data, the system is able to predict conditions 120 hours into the future.

The immediate effect of the disaster was a concerted effort to deploy buoys on Lake Superior. By 1979, eight weather buoys had been placed on the Great Lakes. Today, Lake Superior has 30 buoys and weather stations, which are part of the larger Coastal Marine Automated Network (CMAN) to measure wind direction and speeds, air temperature, and surface pressure.

Still, the buoys are not without their drawbacks. Beginning in November, most of the buoys are pulled to save them from the ravages of ice. As Matt Zika of the NWS explained to the Detroit News, “A lot of times, by the latter part of that storm season, we don’t even have the buoys out there to measure.” Even though the National Oceanic and Atmospheric Administration is pioneering a system of robots to work in these conditions, developing ice-resistant buoys for year-round deployment remains an engineering challenge.

But even with these limitations, captains on Lake Superior have access to an incredible surfeit of data, like real time satellite-fed radar, GPS, wave forecasts, multiple redundant methods of communication, and weather data from the buoys, which would have made all the difference that November night. And the real proof is that the data is relied upon. “When I started here 20 years ago, none of the boat crew trusted any forecast out beyond about 24 hours,” remarked Jay Austin, a professor at the University of Minnesota who works on these models. “Now, if the forecast says it’s going to be bad four days from now, we’re not going out.”

Sailors who once scoffed at forecasts now plan around them. The stoicism embodied by that last message, “We are holding our own,” has been replaced by a quiet confidence in the tools of modern science. The result is that no major Great Lakes freighter has been lost in open waters since the Edmund Fitzgerald sank 50 years ago.

In this way, the Fitz is more than just a shipwreck. It is a boundary marker between eras. It closed the age when human judgment alone had to contend with Great Lakes, and opened another where human intelligence is amplified by machines. The price of that knowledge was 29 lives, and thankfully, it is a legacy that no others have followed.

At the Roots of Progress Conference earlier this month, Tyler Cowen interviewed OpenAI CEO Sam Altman, who at one point wondered about the world that was to come with AI agents. In the not too distant future, he imagined that AI agents would be involved in every aspect of business and would even negotiate with other agents autonomously, only involving their human counterparts when specific guidance was needed. It was a passing comment, but one that captured the scale of institutional change that agents might bring. But it also got me thinking about the legal and practical issues with AI contracting.

A couple of days after the talk, I saw that “Contractibility Design” by Roberto Corrao, Joel P. Flynn, and Karthik Sastry was posted on the National Bureau of Economic Research. Apparently, it has been floating around since 2023 but, truth be told, I probably wouldn’t have given it much attention if Altman’s comment wasn’t brewing in the back of my mind.

The paper deals with a long-standing problem in law and economics known as incomplete contracts. The way to think about incomplete contracts is to start with their antonym, complete contracts. Complete contracts are ideal in that they specify for both parties their rights and duties for every possible future state of the world. In the strictest sense, no contract is ever complete. However, as contracts become more clearly defined, they move further away from being incomplete. It’s also important to note that even if a contract is incomplete, that doesn’t mean it cannot be enforced by courts, and just because a contract is incomplete doesn’t mean it isn’t economically optimal.

Research on incomplete contracts can be traced to work from the criminally underrated economist Herbert Simon, who first modeled the employment relationship as an incomplete contract in his 1951 paper “A Formal Theory of the Employment Relationship.” But it took the work of Oliver Williamson and especially Oliver Hart, Sanford Grossman, and John Moore to flesh out the implications for ownership, control rights, and governance structures. In short, where markets fail to specify everything in advance, contracts often fill in the gaps.

What struck me about Corrao, Flynn, and Sastry’s approach is that they treat incomplete contracts like an optimization problem. In their terminology, coarse contracts are those contracts with low fidelity. Think of it like a low resolution, pixelated video. As you move toward a complete contract, resolution increases. They then use this idea to help understand why “even billion-dollar commercial contracts can be perplexingly imprecise and filled with phrases like ‘best efforts,’ ‘reasonable care,’ and ‘good faith.’” One might expect that with so much money on the line, parties would pay for absolute precision. Yet in practice, vagueness persists, often by design.

Corrao, Flynn, and Sastry show that the fuzziness is not a flaw, but a feature. It’s the result of balancing two opposing costs: front-end costs and back-end costs. Front-end costs, or ex ante costs, involve everything that comes with drafting contracts, foreseeing contingencies, and writing precise language. On the other hand, back-end costs, or ex post costs, come from generating evidence and proving breach after the action occurs.

Their central finding is counterintuitive. Even tiny front-end costs can make coarse, incomplete contracts optimal, even for high-stakes deals. As they write, “Thus front-end costs always lead to incomplete contracts: there are finitely many contingencies, but within each contingency many actions are legally permissible.” Surprisingly, back-end costs, no matter how large, do not cause contract coarseness. In fact, it’s the opposite: Back-end costs lead to complete contracts.

Like so much in economics, the difference comes down to marginal costs and marginal benefits.

Imagine you already have a contract with 10 levels of performance including acceptable, good, very good, and so on. Adding an 11th level by splitting good into good and slightly better than good requires writing rules distinguishing this new level from all the nearby levels above and below it. As you go down this route, the number of distinctions grows rapidly. Mathematically, it’s proportional to n². Still, the benefit is tiny because you’re only fine-tuning decisions for the narrow slice of cases that fall in between “good” and “very good,” which adds little value.

Back-end costs change the calculus entirely. Here, the principal pays only to verify the actions that actually occur, not every hypothetical contingency. Because the design of the contract shapes which actions are likely to arise, the costs and benefits remain roughly aligned. Greater precision pays off when it helps resolve real disputes. Put differently, front-end costs punish precision more than they help, while back-end costs rise in proportion to the value of that precision.

Reading Corrao, Flynn, and Sastry’s paper right after hearing Altman speak got me thinking, what happens when we throw AI agents into the mix?

In theory, large language models could drive down front-end costs. AI systems can already draft contracts, generate contingencies, and reconcile precedents at negligible marginal cost. They could also explore counterfactuals and write clauses for edge cases that humans would never think to include. If Corrao, Flynn, and Sastry are right, this technological shift should push us toward more complete contracts. And this wouldn’t be because humans suddenly became better lawyers, but because AI collapses the cost curve of precision.

Nevertheless, there are reasons to believe this might not happen. Even with low drafting costs, there may be computational limits to specifying all contingencies. Even more important, contracts are enforceable only when you can generate evidence showing the agent violated the terms. Data logs, model weights, or natural-language explanations aren’t likely to satisfy legal standards of proof. Moreover, many AI systems operate in probabilistic rather than deterministic ways, blurring the line between intentional and accidental.

Still, all of this assumes that AI agents are able to write contracts, which is a big if. Every US state prohibits non-lawyers from giving legal advice, drafting legal documents like contracts and representing others in legal matters via Unauthorized Practice of Law (UPL) statutes. Telling your AI agent to negotiate and draft a contract with a supplier is a textbook UPL violation because AI is drafting a legal document and acting on your behalf without being a licensed attorney.

Having two AI agents work together on deal terms is a little more complicated because parties can draft their own contracts without having a lawyer around. Still, I suspect that this contract would probably get thrown out in court because the AI is making legal judgments about contract terms and courts generally interpret UPL broadly.

Indeed, a bunch of cases brought against LegalZoom in the last two decades suggest that legal AI agents will face court scrutiny and probably won’t win. Caroline Shipman’s law review article “Unauthorized Practice of Law Claims Against LegalZoom—Who Do These Lawsuits Protect, and is the Rule Outdated?” offers a great overview of these cases, which are too complicated to review here. But courts have ruled against LegalZoom for answering specific questions about contract terms. I can’t imagine autonomous contract negotiation skirts UPL.

In other words, the very legal doctrines designed to protect consumers from unqualified advice may slow the evolution of more efficient, precise contracting. If AI could truly minimize front-end costs, it might finally fulfill the old dream of near-complete contracting. But for now, that world remains, in Altman’s words, “not too distant.”

Until next time,

🚀 Will

A few months ago, I told you that I was restarting Exformation and working on a bunch of new things for this Substack and at AEI. Part of the reason I’ve been so quiet, with only a few posts in the last couple months, is that I spent a good amount of time this summer in founder mode, thinking about projects, working on papers, and executing ideas. I am excited for all of the new work that I am going to be rolling out soon and into next year.

I will be honest. I haven’t been the best about sending my AEI work through this Substack. So if you want to get caught up on my AI work, I wrote an extended summary for my website and I updated my publications list. Otherwise, here are a couple recent highlights:

• My latest is titled “Procedural Rituals Over Governance Results” and is a reflection on Jen Pahlka’s talk at the Roots of Progress conference. It builds on Nicholas Bagley’s concept of procedure fetish, ultimately coming to the conclusion that: “Our procedural culture doesn’t just make it harder for institutions to function: It also shapes the public’s instinct to say no. The tragedy is that the impulse behind participatory democracy—the desire for inclusion and fairness—has translated into bureaucratic systems that confuse involvement with improvement. Until we rediscover forms of governance that reward results instead of rituals, we’ll keep mistaking the appearance of responsiveness for the real thing.”

• In “Senator Sanders’ AI Report Ignores the Data on AI and Inequality,” I did some digging into Senator Bernie Sanders’ new report which claims that 100 million jobs will be lost in the next ten years due to AI. Beyond the issues with how the AI job loss model was constructed (it simply asked ChatGPT), my biggest concern with the report is that it “reviews some key papers on automation and income inequality, but nowhere does it review the current literature showing that new AI tools are reducing inequality. In Brynjolfsson et al. (2023); Caplin et al. (2024); Choi et al. (2023); Hoffmann et al. (2024); Noy & Zhang (2023); and Hauser & Doshi (2024), advanced AI tools were found to be skill equalizers, raising the performance of those at the bottom in customer support, legal work, and software development, among others. If Sanders was truly concerned with worker inequality, he should be optimistic about AI tools and engaging with the empirical work on this subject.”

• In “The AI Revolution in Property Tax Assessment,” I dove into one of the lesser known applications of AI tools in property tax assessment. As I wrote, “traditional assessment methods face a litany of problems. Valuations can often be inconsistent and municipalities are typically understaffed and resource-constrained.” This is where AI comes in because models can be trained on property characteristics, sales data, and market trends “to address the core challenges of traditional property tax assessment methods.”

Please don’t hesitate to send me ideas for future work and if you don’t mind sharing these emails with colleagues that will help me grow this site. Now on to the good stuff…

Empirical economists have been busy during the last couple of months trying to understand how LLMs are impacting companies and labor markets. To help keep everything straight, I spent some time in April compiling everything I could find into a table. Since then, I have been updating it with high quality papers. It excludes surveys, dataset, and indices because I am working to collect those for another project. But as I find material, the table below will be updated, so please send me what you’ve got!

What began originally as a joke has evolved into a serious metric. There is even a Wikipedia page collecting p(doom) estimates. Geoffrey Hinton, the 2024 Nobel Prize winner in physics for his work on AI, puts the number between 10 and 50 percent. Yann Le Cun, the lead of AI at Meta, has his number at less than 0.01 percent. Even Lina Khan, previous Federal Trade Commission Chair, proffered a guess at around 15 percent. It has become a common shorthand in AI risk discussions, allowing researchers and policymakers to quantify and compare their assessments of AI risk.

But I find all the talk of p(doom) deeply unsatisfying.

Let’s take one of the lowest numbers, p(0.05), or a 5 percent chance of doom. Assuming that the clock starts the moment that artificial intelligence is created and every day is an independent trial, it will only take 20 days for doom to occur. However, if you drastically increase your probability to p(0.95), then the expected value is 0.6, which would mean that by mid afternoon, you’re dead.

But notice how I couched this statement. I assumed that every day is an independent trial. But is every day really an independent trial, or should a trial be in mere minutes?

In practice, many p(doom) estimates tend to be unconditional, that is, doom happens regardless of what else might occur. To me, as someone who was trained in Bayesian econometrics, most p(doom) conversations feel like they are the first 5 minutes of an introductory course where you also learn conditionals, and start exploring if probability A increases, decreases, or perhaps follows a bathtub over time, depending on the probability of B.

The more technically astute will update their p(doom) based on changes in the environment. Safety research, governance decisions, and technical architectures should alter future probabilities. Today’s p(doom) should, in theory, be different from tomorrow’s based on what we learn and do today. But most people don’t actually specify their conditionals. I only rarely hear, “my p(doom) is 15 percent assuming current trajectories, 3 percent if we implement interpretability breakthroughs, or 40 percent if we race toward AGI without safety measures.”

Two years ago, Eliezer Yudkowsky had a conversation with podcaster Dwarkesh Patel that perfectly illustrates this fundamental misunderstanding of how probability should work in AI safety discussions. Yudkowsky just released a new book with coauthor Nate Soares titled If Anyone Builds It, Everyone Dies last week, and not surprisingly, he has a p(doom) that is above 95 percent. On X, Jack Rabuck summarized the conversation, saying,

I listened to the whole 4 hour Lunar Society interview with [Eliezer Yudkowsky…] that was mostly about AI alignment and I think I identified a point of confusion/disagreement that is pretty common in the area and is rarely fleshed out:

Dwarkesh repeatedly referred to the conclusion that AI is likely to kill humanity as “wild.”

Wild seems to me to pack two concepts together, ‘bad’ and ‘complex.’ And when I say complex, I mean in the sense of the Fermi equation where you have an end point (dead humanity) that relies on a series of links in a chain and if you break any of those links, the end state doesn’t occur.

It seems to me that Eliezer believes this end state is not wild (at least not in the complex sense), but very simple. He thinks many (most) paths converge to this end state.

That leads to a misunderstanding of sorts. Dwarkesh pushes Eliezer to give some predictions based on the line of reasoning that he uses to predict that end point, but since the end point is very simple and is a convergence, Eliezer correctly says that being able to reason to that end point does not give any predictive power about the particular path that will be taken in this universe to reach that end point.

Dwarkesh is thinking about the end of humanity as a causal chain with many links and if any of them are broken it means humans will continue on, while Eliezer thinks of the continuity of humanity (in the face of AGI) as a causal chain with many links and if any of them are broken it means humanity ends. Or perhaps more discretely, Eliezer thinks there are a few very hard things which humanity could do to continue in the face of AI, and absent one of those occurring, the end is a matter of when, not if, and the when is much closer than most other people think.

In a review of the new book in Wired, Steven Levy asked Yudkowsky about how he expected to die and he retorted by saying, “If you want a more accessible version, something about the size of a mosquito or maybe a dust mite landed on the back of my neck, and that’s that.” Levy continued,

The technicalities of his imagined fatal blow delivered by an AI-powered dust mite are inexplicable, and Yudowsky doesn’t think it’s worth the trouble to figure out how that would work. He probably couldn’t understand it anyway. Part of the book’s central argument is that superintelligence will come up with scientific stuff that we can’t comprehend any more than cave people could imagine microprocessors.

None of this is satisfying to me and it’s my main gripe with the book, which admittedly, I am about halfway through.

The problem with Yudkowsky’s approach isn’t that he’s wrong about AI being potentially dangerous. I, too, am worried about advanced capabilities being harmful. Rather, what irks me is that his framework makes productive discourse nearly impossible. When he makes strong claims that doom is both inevitable and incomprehensible, he puts his argument beyond the reach of rational debate.

The dust mite scenario perfectly encapsulates this problem. By invoking incomprehensible superintelligence, Yudkowsky sidesteps the hard work of modeling how AI systems might cause harm and what we might do about it. Real progress in AI safety requires grappling with concrete failures, understanding specific vulnerabilities, and developing targeted interventions. It’s not as though Yudkowsky hasn’t done this work before, but that’s not what the book is focused on.

Rather than asking, “What’s your p(doom)?” we should be asking, “Under what conditions does risk increase or decrease?” and “Which interventions have the highest expected value for reducing catastrophic outcomes?” This isn’t just an academic exercise. It’s the difference between paralyzing fatalism and productive risk management.

But AB 1018's definitions cast such a wide net, that they would regulate virtually any computational process used in business operations. Even California's State Water Board warned that Excel workbooks could trigger regulatory requirements.

Using four large language models to estimate compliance costs, individual firms could face between $2 million and $6 million over a decade. Even assuming only 5 percent of businesses comply, the total ten-year cost for the entire economy could reach the billions to low trillions. To be fair, these estimates vary widely in their ranges, suggesting substantial uncertainty, and the economy-wide numbers are constructed such that more companies are included that are likely to be regulated.

Still, the compliance cost estimates for AB 1018 reveal a broader pattern emerging across proposed AI regulation. There is an enormous hidden price tag of mandating human oversight in automated decision-making systems. While human oversight in high-stakes decisions has clear value, at what point do the costs of mandating it across entire sectors outweigh the benefits? The rough estimates for AB 1018 suggest we may already be approaching that threshold, and they serve as a cautionary tale for lawmakers considering similar sweeping mandates elsewhere.

Exformation

Tech policy from an outsider, inside the beltway.

By Will Rinehart

AB 1018 Automated Decision Systems

In May, the California Privacy Protection Agency (CPPA) finalized regulations addressing the use of automated decision-making technology (ADMT) under authority granted by the California Consumer Privacy Act (CCPA). Now, the California Legislature is weighing AB 1018, which would significantly expand this regulatory framework by addressing a broad spectrum of covered automated decision systems (ADS).

Under AB 1018, any process "derived from machine learning, statistical modeling, data analytics, or artificial intelligence that issues simplified output, including a score, classification, or recommendation, that is designed or used to assist or replace human discretionary decisionmaking and materially impacts natural persons" would face regulatory oversight.

AB 1018 casts a notably wider net than the CCPA regulations in defining consequential decisions. It would apply to any decision that impacts the cost, terms, quality, or accessibility for employment-related decisions; education and vocational training; housing and lodging; anything that involves your utilities; family planning, adoption services, and reproductive services; health care and health insurance; financial services; the criminal justice system; legal services; arbitration; mediation; elections, access to government benefits or services; places of public accommodation; insurance; and internet and telecommunications access.

The proposed rules distinguish between the developer—the groups that produce an automated decision system—and the deployer—or the entities that use a covered ADS to make or facilitate a consequential decision.

To be compliant, developers will need to

• Conduct impact assessments before deployment and annually thereafter;

• Provide detailed documentation to deployers about system capabilities and limitations;

• Submit to third-party audits (starting in 2030); and

• Designate compliance officers.

For deployers, the rules would require:

• Transparency: Provide clear disclosures to people subject to automated decisions, explaining what system is being used and how;

• Opt-out rights: Allow people to request human decision-making instead (with some exceptions for financial services and medical emergencies);

• Appeal process: Give people the right to correct errors in their data and appeal adverse decisions; and

• Impact assessments: Large-scale deployers must conduct regular third-party audits.

Government Estimates of the Cost of AB 1018

Whatever human-centered benefits AB 1018 might deliver, the price tag will be staggering. California's fiscal analysis shows the bill would impose significant costs on nearly all state agencies: "In the aggregate, for all affected state agencies, costs may be in the hundreds of millions of dollars annually ongoing." These estimates reflect substantial financial impacts across virtually every sector of state government, from individual departments facing millions in compliance costs to the judicial system potentially requiring hundreds of millions annually.

For example:

• The State Controller’s Office (SCO) estimated compliance at $3,553,000 from 2025 to 2028, which would include “5–7 information security staff—including a dedicated supervisor—for the initial implementation of AB 1018, with 3–4 (including the supervisor) required for ongoing security and compliance operations related to automated decision systems.”

• The Department of Justice (DOJ) estimates a fiscal impact of $2.5 million or less.

• The Civil Rights Department (CRD) estimates a fiscal impact over $2 million.

• The State Water Board estimates a significant fiscal impact (General Fund, special funds) likely in the millions of dollars per year. The State Water Board reports, as written, this bill’s definition of automated decision system is vague, ambiguous, and could encompass many current tools used, like Excel workbooks. These tools are used broadly across Water Boards programs, and many are used to inform actions that could be considered consequential actions under the bill.

• “UC Health estimates a fiscal impact of $42 million in one-time costs and approximately $24 million in ongoing costs to implement AB 1018.”

• “The California Department of Health and Human Services (CalHHS) reports a significant fiscal impact (General Fund, special funds)...in the low millions up to low tens of millions (Total Fund) annually as solutions applicable under the bill are expected to grow in use over the coming years.”

• “Unknown, potentially significant costs to the Judicial Council, likely in the hundreds of millions annually.”

AB 1018 extends far beyond government. It covers countless industries: banks and credit unions; mortgage companies and brokers; insurance companies; investment firms and financial advisors; credit card companies; Software-as-a-Service (SaaS) providers; social media platforms; E-commerce platforms; HR technology vendors; educational technology companies; hospitals and health systems; health insurance companies; medical device manufacturers with AI components; dental and vision care providers; staffing and recruiting agencies; gig economy platforms (Uber, DoorDash, etc.); companies using AI for hiring, performance reviews, or scheduling; property management companies; real estate platforms and brokerages; home appraisal companies; public housing authorities; educational institutions (universities, colleges, K-12 schools); student loan servicers; educational assessment companies; electric utilities; water companies; telecommunications providers; waste management companies; law firms using AI tools; legal technology companies; arbitration services; mediation services; online marketplaces; retailers using AI for pricing or recommendations; hotels, restaurants, entertainment venues; as well as government contractors; and benefits administration companies.

The sheer breadth of industries that would fall under AB 1018's umbrella creates a unique challenge for policymakers: How do you estimate the economic impact of a bill that could affect virtually every sector of the economy?

AB 1018 Cost Estimates

I've been experimenting with large language models (LLMs) to estimate AI bill compliance costs. In March, I published “How much might AI legislation cost in the U.S.?,” which explained the pluses and minuses with this method. What was surprising is that the LLMs were usually close to first year estimates, but they tended to predict much higher ongoing annual costs, suggesting a systematic underestimation.

In June, I published a piece in the City Journal on New York’s RAISE Act which I subjected to the same kind of method. As I wrote,

I asked the leading LLMs to read the RAISE Act and estimate the hours needed to comply with the law in the first year and in every year after that for a frontier model company. The results, displayed in the table below, suggest that initial compliance might fall between 1,070 and 2,810 hours—effectively requiring a full-time employee. For subsequent years, however, the ongoing burden was projected to be substantially lower across all models, ranging from 280 to 1,600 hours annually.

The wide range in estimates underscores the fundamental uncertainty with the RAISE Act and other similar bills. The fact that sophisticated AI models are not converging on consistent compliance costs suggests just how unpredictable this legislation could prove in practice. The market is moving quickly. We need laws that prioritize effective risk mitigation over regulatory theater.

During a Hackathon in May, I coded up a first version of a prompt script, which I am going to iterate on in the future. It is built on persona prompting, where you direct the LLM to take on a role and then answer questions. So, I set up the script to reflect different industries, resources, legal teams, and familiarity with the law.

The benefit of this kind of scripting is that you can run many estimates simultaneously and then summarize the results. In the future, I am going to match these personas with what we know about the market to create more accurate predictions. The code is still messy, but I intend to come back to it when I work on the full paper this fall.

To continue this work, I thought I would try to push the LLMs further in my analysis of AB 1018. What’s especially interesting, and worrying about this law is that it will apply to so many businesses. Could LLMs help figure out the scope of the law?

I followed the typical method of running a cost calculation in public policy. First, you estimate the hours of compliance, then multiply it against market labor rates to calculate an economic cost for a firm. Then you take this number and multiply it against the number of impacted businesses. However, estimating the number of impacted businesses tends to be a blunt measure, so I used a second set of scripts to then estimate which industries are likely to be affected by the law. All of the data can be found in this spreadsheet.

I tested four models, Claude, ChatGPT, Gemini, and Grok, using the July 17, 2025 version of AB 1018. Complete chat transcripts are available here, with notes where calculations were incorrect. The consolidated results of the labor hours from all four LLMs are summarized in the table below:

A couple items worth noting from these estimates:

• ChatGPT had low to moderate estimates across the board. The costs for developers jump in 2030 with audits, but still stay below Claude’s and Grok’s numbers. Like Claude and Gemini, deployers are estimated as less burdened than developers.

• Claude had the highest estimates among all LLMs, projecting substantial compliance burdens for both developers and deployers. Developers especially face steep increases after 2030 audits.

• Gemini has the lowest estimates across all categories, and are less than half of ChatGPT’s and a fraction of Claude’s.

• Grok originally estimated cost based on two ADS’s, so these numbers were adjusted to just one ADS. Given this change, Grok provides middle-of-the-road compliance cost for developers, but the highest burden for deployers. Grok stands out by flipping the assumption: It estimates deployers will face heavier costs than developers, especially on an ongoing basis.

It is worth noting that, without prompting, each of the LLMs broke out the labor costs to developers into three buckets: the first year of compliance, an annual cost from 2027 to 2030, and an annual cost from 2030 onward, which is when third party audits are slated to begin. However, Gemini didn’t give an estimate for 2030 onward, saying, “This will add significant hours and cost, involving contracting with an auditor, providing necessary information (with redactions for trade secrets), and making a high-level summary of feedback publicly available. The time commitment for this is highly variable.”

Taking it one step further, I then asked the LLMs to translate the labor compliance hours into costs using typical labor wage rates, allowing the LLMs to select their own to see how that would vary. Each of the LLMs selected a slightly different method:

• Gemini assumed a standard 50/50 blended rate: “A blended rate of $150 to $250 per hour is used to account for the mix of senior technical and legal staff required for these tasks.”

• ChaGPT varied this blended ratio: “Legal/Compliance staff: $150/hr; Technical staff (engineers, data scientists, IT): $100/hr; Roughly 60% of work = technical, 40% = legal/compliance (developers skew a bit more technical, deployers a bit more compliance-heavy).”

• Claude offered a range of wage assumptions, but didn’t explain which ratio it selected: “Senior Legal/Compliance Officer: $150-200/hour; Mid-level Compliance Analyst: $75-100/hour; Technical Documentation Specialist: $60-80/hour; Administrative/Coordination: $40-60/hour; Blended average for compliance work: ~$85-110/hour”

• Grok had the clearest breakdown: “~40% data scientist ($80/hour for evaluations, ~216-356 hours), ~30% compliance officer ($70/hour for oversight/setup, ~162-267 hours), ~20% legal ($115/hour for reviews, ~108-178 hours), ~10% admin ($35/hour for records, ~54-89 hours).” Still, these don’t sum to the $70/hr and $65/hr that was used in the final cost calculation.

The variation in wage rates is charted below.

Combining the labor hours with the wage rates yielded the estimated yearly and annual costs, which are displayed below.

Looking at these results, some trends emerge:

• ChatGPT produced moderate estimates overall. Still, the estimates suggest developers face higher burdens than deployers, especially post-2030.

• Claude offers the highest potential cost since the first-year developer cost could be $132,000. Both developers and deployers face very large costs, with developers’ post-2030 costs nearly doubling their annual burden.

• Gemini shows the widest spreads as developer first-year ranges from $27K to $105K. It consistently gives the lowest annual ongoing costs, especially for deployers. Similar to the previous analysis, Gemini makes compliance look far less expensive long-term.

• Grok sits in the middle-to-low range for developers, but suggests costs will be higher for deployers. It reinforces Grok’s distinctive assumption: Deployers, not developers, bear the heavier financial load.

From here, I projected these costs over a ten-year period and applied standard economic discounting methods, which I explained back in March:

Similar to discounted cash flow (DCF) analysis, we can think of regulation as a stream of future costs that a regulation is expected to generate over the 10 year period. By summing them all up and then discounting them back to present value using a discount rate that reflects the costs’ long-term time horizon, we can estimate the current market value of the regulation.

Estimates for a ten-year discounted regulatory costs (DRC) are collected below for each of the regulation types and LLMS:

These ten-year totals, which are common in cost benefit analysis, represent the present value of sustained compliance costs that firms would need to factor into their long-term business planning and automated system adoption decisions. The estimates span a 13.4-fold difference, from Gemini's most conservative deployer projection of $620,000 to Claude's highest developer estimate of $8.3 million. Most projections cluster between a $2 to $6 million range for the decade-long compliance period.

Still, the substantial range in estimates reflects the fundamental uncertainty about how AB 1018's broad definitions would be interpreted and implemented in practice.

NAICS Matching and Discounted Regulatory Costs

To translate the LLM-generated industry classifications into quantifiable economic impacts, I developed a systematic approach to match potentially affected industries with California's official industry statistics.

I asked all of the LLMs to create a list of businesses that would likely be regulated by the bill and then I had them match those guesses with the corresponding NAICS code. The responses from ChatGPT, Claude, Grok, and Gemini are collected in this Google Doc.

Each LLM produced different approaches to industry identification. ChatGPT generated 65 business classifications, Claude produced 92, Gemini identified 19, and Grok classified 30 distinct business types. These classifications varied significantly in their specificity. Some LLMs provided detailed 6-digit NAICS codes (like 325412 for pharmaceutical companies), while others used broader industry ranges (like Gemini's "31-33" for manufacturing).

The fundamental problem lies in matching all of these choices with the limited California data. The most up-to-date public business information for California is provided every quarter on a delay. More importantly, these numbers don’t offer the specificity to match businesses with their industry classification. So, I used a hierarchical matching function to place the regulation in the most specific bin possible, while retaining whether the businesses would face deployer, developer, or both regulations. The code helped systemically map LLM-generated regulatory predictions and official state business statistics. This approach successfully matched 96 of the 98 LLM classifications (98 percent match rate) to California establishment data. Match data is located here.

Finally, I took the cost estimates from the section above for each company and then multiplied them against the number of businesses in each category as well as the regulatory type (developer, deployer, or both). If an industry faced both types of regulation, I just added together the two costs. The total discounted regulatory costs are collected below.

I also reran the estimates assuming that only 5 percent of companies will need to comply. It is a blunt measure that I intend to alter in future work. But even with that low number, as the table below shows, the costs may very well be in the trillions.

I don’t want to dwell too much on these numbers because they are rough estimates. The LLM classifications represent informed predictions rather than definitive legal interpretations of AB 1018's scope. More importantly, the NAICS matching process necessarily involves aggregation. Specific business types identified by LLMs are matched to broader industry categories in the official data. This means that the estimates are sure to be higher than actual impact.

Despite these limitations, this approach provides the first systematic attempt to quantify AB 1018's potential economic scope using computational methods, offering policymakers concrete estimates to anchor discussions about regulatory costs and benefits.

Commentary on All The Cost Estimates

Unlike targeted regulations, AB 1018 forces regulatory convergence across disparate sectors. Healthcare AI systems, bank credit algorithms, retail pricing models, and utility grid management would all face similar compliance requirements despite vastly different contexts and risk profiles.

This one-size-fits-all approach ignores the reality that automated systems in different industries serve different functions, carry different risks, and require different safeguards. A scheduling algorithm at a restaurant poses fundamentally different risks than an AI system screening loan applications, yet AB 1018 would subject both to similar regulatory frameworks.

This approach effectively criminalizes mathematical efficiency. Businesses that spent years optimizing operations through data-driven automation face an impossible choice. They could either abandon efficiency gains or navigate complex compliance regimes that could cost millions per system. Neither are especially palatable.

Still, these rough estimates represent just the tip of the iceberg. They capture only the direct compliance costs of hiring staff, conducting audits, implementing new processes, and maintaining documentation. What they don't account for are the cascading economic effects that would ripple through entire sectors. Every dollar spent on regulatory overhead is a dollar not invested in innovation, service improvements, or competitive pricing. For the economy as a whole, it would represent a massive shift of resources from productive activities to regulatory compliance. When California's State Water Board warns that the bill's definitions are so broad that Excel workbooks could trigger regulatory requirements, we're looking at a fundamental restructuring of how modern organizations operate.

Consider the practical implications. A community bank using automated credit scoring to expand lending to underserved populations might abandon the technology rather than spend $3 million on compliance infrastructure. The result isn't more human oversight; it's less access to credit. A healthcare system using AI to prioritize patient triage might revert to manual processes that are slower and more error-prone, not because human judgment is superior in this context, but because the regulatory burden makes the technology economically unviable.

Start-ups and smaller companies face particularly acute challenges. While established corporations can spread compliance costs across large revenue bases, emerging businesses may find regulatory requirements consume resources that would otherwise fund product development, hiring, or market expansion.

The choice isn't between perfect algorithms and perfect human judgment. Neither exists. The choice is between regulatory frameworks that preserve beneficial innovation while addressing legitimate risks, and those that impose such crushing compliance burdens that they eliminate both the risks and the benefits of technological progress.

Despite its human-centric rhetoric, AB 1018 may ultimately prove deeply inhumane to the Californians it claims to protect.

We need to get ahead of this thing.

I've heard this refrain countless times over the past two years in AI policy circles. But listen closely, and you'll discover the real message underneath, a quiet admission of past failure: We didn’t move quickly enough when it came to social media and we cannot make the same mistake with AI.

Just this year, over 1,000 AI bills have been proposed in the states. Thankfully, some of the worst have been stopped, and for the most part, state sessions are done for the year. But next year, legislators will be back at it.

The breakneck speed at which states are rushing to pass AI legislation has been deeply troubling, largely because the quality of these bills has been so poor. Colorado's experience is illustrative of this. Despite being the first state to pass comprehensive AI legislation, lawmakers have been debating a major overhaul of it due to the law's flawed construction. Neil Chilson, former chief technologist at the FTC and current Head of AI Policy at the Abundance Institute, recently pointed out just how absurd the bill is, saying:

Consider Colorado’s SB 24-205, a “comprehensive” AI law passed last year, which casts such a wide net that legislators clarify in the text that calculators and spellcheckers don’t count as AI … unless they become “a substantial factor in making a consequential decision.” How’s that for clarity?

Even more telling is the European Union's recent hesitation. After years of positioning itself as the global leader in AI regulation, the EU is now considering delaying implementation of its own AI Act due to practical concerns about how the law would work in practice. When both pioneering state legislators and seasoned international regulators start having second thoughts about their own AI laws, it suggests that the rush to regulate may be outpacing our understanding of how to regulate effectively.

So I was happy to see the House approve a budget bill that contained a 10-year moratorium on state AI regulation. While it's probably longer than optimal, Congress has needed to assert its authority in this space.

The way that the bill is written, states could still use existing laws to address AI issues. Privacy laws, consumer protection rules, and fraud statutes still apply to AI companies. The moratorium doesn’t completely preempt state law, it simply stops states from creating new, AI-only regulations for a while, as Congress figures out a national approach.

The inclusion of the provision kicked off a big debate about the merits and demerits of an AI moratorium. As one would expect, a lot of state legislators aren’t a fan. But in an unexpected twist, Colorado Governor Jared Polis, who actually signed his state's AI Act into law, backed the federal moratorium. According to a statement from his office, Polis believes "a strong national policy governing AI consumer protection that supersedes state law would be the best course of action, as a state-by-state patchwork creates a challenging regulatory environment and would leave consumers worse off overall." The governor supports suspending state AI laws for several years, giving Congress the breathing room needed to craft what he calls "a true 50-state solution to smart AI protections for consumers while driving innovation." It's remarkable for the same official who enacted the first state-level AI regulation to now argue for hitting the pause button.

Policy guru Adam Thierer proposed the AI moratorium concept last year, drawing from his extensive experience in AI policy. Having witnessed the regulatory chaos surrounding privacy legislation, he cogently argued that a similar scenario was now unfolding with artificial intelligence. Since California passed the Consumer Privacy Act in 2018, states have rushed to enact their own privacy laws. Today, 19 different state privacy bills exist, each with their own unique requirements and structures. This patchwork approach forces businesses to constantly adapt their compliance strategies with every new law that passes, creating exactly the kind of regulatory mess that Congress should have addressed years ago. But the federal gridlock persists largely because Senator Maria Cantwell, who chairs the key Senate committee, opposes any federal framework that would supersede California's state law. I explored this dynamic in detail last year.

The regulatory burden for privacy has become so complex that a senior policy executive at a major tech company once told me something I’ll never forget: We will be [legally] defensible, but I am not sure we could ever be technically compliant. This stark admission reveals why Thierer's call for an AI moratorium deserves serious consideration. We are on track to repeat the same mistakes with artificial intelligence that we made with privacy regulation.

The strongest argument I have heard against the moratorium is that it takes options off the table. I have written a lot about regulatory options, which you can find here and here. Optionality, like so much else, is a double-edged sword. Benefits on one side of the ledger appear as costs on another . The option to regulate for legislators often appears as a drag on investment. The moratorium flips this equation. By temporarily removing the option to regulate at the state level, it trades legislative flexibility for market certainty. Companies would know that for a defined period, they won't face a constantly shifting landscape of state-by-state requirements.

To be fair to the critics of the moratorium, a decade is a significant stretch of time, and AI's potential for disruption is undeniable. The technology is advancing rapidly, and legitimate concerns exist about leaving it completely unregulated for too long. That's why I'd be perfectly content with a five-year moratorium. Indeed, five years might be the sweet spot. It would give Congress enough time to study the technology, understand its implications, and craft thoughtful federal legislation without the pressure of competing state laws proliferating in the background. This is why I'm giving the current moratorium idea two cheers rather than three.

The predictable objection to any federal moratorium will be that Congress moves at glacial speed. But this frustration, while understandable, misses the larger constitutional point. When businesses operate across state lines, as all major AI companies do, a patchwork of conflicting state regulations creates exactly the kind of interstate commerce problem the Constitution empowers Congress to address. Right now, lawmakers can point to state activity as evidence that the system is working and avoid making difficult decisions. Remove that escape hatch, and suddenly the pressure to craft thoughtful federal legislation becomes much more intense. And the pressure should be squarely on federal lawmakers to act. Not just on AI, but also on the broader suite of technology issues they've been dodging for years.

Still, the moratorium is a pragmatic compromise that prioritizes getting regulation right over getting it fast. We need smart AI regulation. But we need it to be consistent and evidence-based. A temporary pause on state-level rules gives us the best chance to get this right.

Policymakers are rushing to regulate artificial intelligence (AI), but the economic impact of these regulations remains largely unexplored. While the European Union and the United Kingdom have produced cost estimates, recent developments in the United States offer important new benchmarks. Recent amendments to the California Consumer Privacy Act (CCPA) and regulations implementing President Biden’s Executive Order on AI offer crucial insights into what businesses might expect to pay for compliance. The financial burden could be substantial, running into billions of dollars across the economy. Especially as states push to adopt AI bills, understanding these costs is essential for crafting regulations that balance innovation, safety, and economic viability.

Still, these compliance cost estimates are notoriously unreliable. As an alternative approach, I tested whether large language models (LLMs) could provide more realistic estimates by simulating compliance scenarios in the final section of this post. I prompted ChatGPT, Claude, and Grok to act as compliance officers at companies subject to new CCPA provisions and a Bureau of Industry and Security (BIS) rule, asking each to estimate hours needed for first-year implementation and ongoing compliance.

The big takeaways:

• For California's risk assessment regulation, Claude and Grok project 400-580 hours will be needed for the first-year of compliance (vs. the official 120 hours) and 150-240 hours thereafter (vs. the official 18-36 hours annually). ChatGPT estimates the time at 90-250 hours initially and 40-150 hours for each additional year.

• For the automated decision-making provision of the CCPA, Claude and Grok project 450-730 hours for first-year compliance, far exceeding the official 360-hour estimate. While ChatGPT suggests lower initial costs (80-300 hours), all three LLMs predict significantly higher ongoing annual costs than official projections.

• For the BIS reporting rule, Claude projects 1,140 hours for first-year compliance (3x the official estimate of 333 hours), while ChatGPT estimates 280 hours and Grok projects 380 hours. All three LLMs agree ongoing compliance will require substantial resources.

In short, the LLMs consistently predicted much higher compliance costs than the official sources, suggesting a systematic underestimation.

The following analysis unfolds in three parts: The first section examines the compliance costs of California's new regulations. The second section analyzes the BIS's proposed AI reporting requirements. The final section tests whether large language models can provide more realistic cost estimates than traditional regulatory impact assessments.

While these AI-generated estimates aren’t perfect, they suggest that LLMs might eventually serve as virtual simulation environments that capture more diverse business experiences than traditional estimation methods. I’m hopeful that they could bring much-needed empirical grounding to regulatory cost assessment.

The costs of CCPA’s new provisions

In September 2024, Governor Gavin Newsom of California signed into law three bills that amended the California Consumer Privacy Act of 2018 (CCPA). Among other changes, California now requires companies that are governed under the CCPA to prepare a cybersecurity risk assessment (RA) as well as give consumers the ability to opt-out of automated decision making technology (ADMT). As part of the rulemaking, California commissioned a standardized regulatory impact assessment (SRIA), which was prepared by economists at Berkeley Economic Advising and Research and California State University, to estimate the cost of the new regulations. The following analysis examines these official government estimates, which were created through traditional economic modeling methods, not using LLMs. (The LLM-based cost estimation approach is explored later in section 3 of this post.) Since the RA and ADMT provisions have been included in other AI bills, including Colorado’s AI bill and a proposal in Connecticut, the assessment offers a lower bound for what AI regulation might cost in other settings.

The risk assessment regulations require businesses to conduct risk assessments when their processing of personal information presents significant risk to a consumers’ privacy. This includes businesses that sell or share personal information (PI), process sensitive PI, or use ADMT for significant decisions. As part of these new requirements, businesses will need to explain the purpose of processing and the categories of information processed. They will also need to stop their processing if risks outweigh benefits. To prove their compliance, businesses will now be required to submit certifications and abridged versions of their risk assessments to the state government annually.

The rules for ADMT are also involved. Businesses will need to inform consumers about their use of ADMT with a Pre-use Notice and then give consumers a chance to opt-out when ADMT is used for significant decisions. This includes access to financial services, housing, education, and employment; extensive profiling; or training uses of ADMT.

The SRIA's calculations for the cost of the two provisions were based on what fraction of all CCPA-covered businesses would also be subject to these specific requirements, yielding three scenarios:

• Low scenario: 25% of all CCPA-covered businesses would be subject to these provisions (13,082 out of 52,326 businesses)

• Medium scenario: 50% of all CCPA-covered businesses would be subject to these provisions (26,163 out of 52,326 businesses)

• High scenario: 100% of all CCPA-covered businesses would be subject to these provisions (all 52,326 CCPA-covered businesses)

After talking with people in the industry, the SRIA’s economists estimated that 120 hours would be needed to complete the organizational and regulatory requirements for the RA for the first year for each firm. To determine an appropriate wage rate, they analyzed the different types of professionals who would likely be involved in implementing the regulations, resulting in a blended hourly rate of $56.40, based on “an hourly wage of $25.91 for office and administrative support occupations, $42.67 for compliance officers, and $100.61 for lawyers.”

Total first-year costs were calculated by multiplying the per-business cost by the number of businesses in each scenario, for a low range of $89 million, a medium range of $177 million, and a high range of $354 million.

For ongoing costs in subsequent years, the average of these three one-year costs, which was roughly $207 million, was then used to calculate a low range and a high range. As was explained in the SRIA, “We estimate that subsequent RAs—including both new RAs for new PI-processing, as well as reviews and updates to existing PI-processing—will represent 15–30% of total year one compliance costs, with the higher compliance cost threshold occurring in earlier years before gradually falling.”

The table below captures the first year estimates and ongoing costs.

The SRIA is a little more opaque about the paperwork cost for ADMT. The assessment notes that the “compliance estimates come from our own understanding of the regulatory delta and how it relates to the other hourly estimates we derived from industry experts on the RA and CSA regulations.” This part of the assessment assumes that programmers will need between 240 hours and 360 hours to complete the changes. Assuming the software developer hourly rate of $91.14, this corresponds to $21,874 to $32,810 in first-year direct compliance costs per business. Much like the RA regulations, the study estimates “that subsequent years will represent 15% – 30% of total year one compliance costs, with the higher compliance cost threshold occurring in earlier years before gradually falling.”

Putting it all together,

“First-year total costs are comprised of approximately $369M in costs associated with updates to CCPA regulations, $2.0B in costs associated with CSA, $207M in costs associated with RA, and $835M in costs associated with ADMT. While CCPA updates do not have estimated ongoing costs, there are ongoing annual costs associated with each of other elements including CSA (estimated range of $308M-$615M per year), RA (estimated range of $31M–$62M per year), and ADMT (estimated range of $125-$250M per year).”

To be honest, I’m not a huge fan of how this study averaged the low, medium, and high scenarios for business compliance. While it is common to game out high, medium and low scenarios in these analyses and do averaging, it is critical that we estimate a best guess upper bound, as well as a best guess middle bound, and lower bound. If a rule is meant to apply to every business, then our top line estimate should reflect that assumption.

To be clear, this doesn’t mean that we shouldn’t vary the compliance costs by firm. In fact, the estimate for ADMT does exactly this, varying the number of hours needed to complete the requirement between 240 hours and 360 hours. Rather, it means we need to take this 360-hour estimate and compute its absolute top end cost through all calculations.

I would go further and vary the wage rate as well. Every SRIA estimate assumes that the work will be done in-house but many small businesses will need to contract out the work to specialized firms. These costs won’t be a fraction of an employee’s time but an ongoing yearly contract.

Contractrates.fyi, a site that crowdsources contract data, puts the cost of cybersecurity consultants at $148 per hour, almost triple the cost of the SRIA for the risk assessment part. This would mean that businesses would have to spend $17,760 to be compliant, not $6,768. Similarly, the market rate for frontend engineers comes to $395 per hour, which is over four times the official rate for programmers. Cost estimates for full compliance of all business at a market rate are tabulated below. They still include low and high scenarios, which assume that ongoing yearly costs might be either 15% or 30% of the average cost of the first year.

While the official study does use these numbers as inputs in a macroeconomic model, they don’t estimate the total cost of the rules in current dollars. Similar to discounted cash flow (DCF) analysis, we can think of regulation as a stream of future costs that a regulation is expected to generate over the 10 year period. By summing them all up and then discounting them back to present value using a discount rate that reflects the costs’ long-term time horizon, we can estimate the current market value of the regulation. The table below calculates those regulatory costs using the federal two percent discount rate for three scenarios.

Using the SRIA’s numbers, the ten year discounted cost for the risk assessment provision is currently valued between $4 billion and $7 billion, while the cost for the ADMT rule is between $18 billion and $28 billion. But those numbers jump if we assume that every business needs to be compliant. In this case, the RA is valued between $8 billion to $12 billion, and the ADMT could mean $31 billion to $49 billion in current costs. In the most extreme case, the market rate for the cost of RA compliance could be $32 billion, while ADMT rules might be currently valued at $210 billion.

Truth be told, my preferred cost estimate is the middle tranche, the one that models full compliance using the official wage rates rather than market rates. This approach acknowledges that all covered businesses will eventually need to comply, without making the extreme assumption that all will need to hire outside consultants at premium rates.

Admittedly, those are eye popping numbers, but I think it reflects what’s being proposed. If regulators truly want to mandate a comprehensive AI governance regime across the entire economy, we should expect substantial costs. None of this is to say that the costs might not be justified. But policymakers need to grapple with a realistic estimate of economic burden, especially given the uncertain nature of AI risks and the potential for compliance costs to disproportionately impact smaller innovators.

Proposed Department of Commerce regulations

Besides the newest CCPA provisions, proposed rules from the Bureau of Industry and Security (BIS) offer another look at how much AI legislation might cost. The proposed rules, which were part of Biden’s Executive Order on AI, would require covered U.S. persons with advanced AI models or clusters exceeding the technical thresholds to notify the BIS on a quarterly basis. It would also require covered persons to respond to follow-up inquiries by BIS within 30 days.

As noted in the official filing, “BIS estimates the specific survey required by this proposed rule will have an estimated burden of 5,000 hours per year aggregated across all new respondents,” which the agency estimated was “between zero and 15 companies” at the time of writing. Assuming that 15 companies would have to shoulder those 5000 hours, each company would be burdened with 333 hours of compliance, similar to what California estimates for risk assessments.

However, in docket filings for the rulemaking, the US Chamber, TechNet, and the Computer and Communications Industry Association all said that the 5,000 hours significantly underestimated the actual burden of quarterly reporting. Peko Wan, CEO at Pundi X, a company that enables retail stores to accept crypto payment, also thinks the number is too low. She told the trade publication PYMNTS that companies could be burdened between $570,000 and $815,500 annually to comply with the new mandates.

Estimates of paperwork burden hours are known to have problems. In 2018, the GAO reviewed 200 estimates to find that 76 of them did not convert burden hours into dollar costs, which was also the case for the BIS rule. Importantly, the report “found that one [information collection request] underestimated burden by as much as $270 million, and another overestimated burden time by more than 12 million hours.” Internal review processes failed to catch significant errors.

Agencies often lack standardized, data-driven methods for estimating paperwork burden and instead rely on the personal judgment of program analysts. As a result, burden estimates are “not based on objective, rigorous, or internally consistent methodologies” in many cases​. Officials themselves admit skepticism about their own numbers, and in practice, there is little process to verify or validate an estimate once a collection is underway.

LLM estimate of costs

LLMs might offer a pathway towards better estimates. I’ve been especially interested in a recent paper from Park et. al. (2024) that “simulates the attitudes and behaviors of 1,052 real individuals—applying large language models to qualitative interviews about their lives, then measuring how well these agents replicate the attitudes and behaviors of the individuals that they represent.”

This approach could be adapted to simulate how individuals and businesses interact with regulatory paperwork, capturing a more diverse range of experiences than traditional estimation methods. By modeling various user personas that differ by industry, LLMs could generate more nuanced time and cost estimates that reflect real-world variability. While not a complete solution, this method could complement existing approaches by introducing more consistency and empirical grounding to what has largely remained a subjective process.

As a first test of the idea, I prompted ChatGPT, Claude, and Grok to act as a compliance officer at a company being regulated by the two CCPA provisions and the BIS rule, and then asked each LLM to estimate the amount of hours needed to comply in the first year and every year after that. The full chats are linked at the end of this post. The three estimates, which include the RA provisions, the ADMT provisions, and the BIS rule, are captured below, along with the official estimates.

The data reveals striking disparities between official government estimates and those generated by large language models, suggesting that agencies may be systematically underestimating compliance burdens for new AI regulations.

Some things I noticed:

• Estimates for California’s risk assessment (RA) regulation vary widely. Both Claude and Grok estimate that 400-580 hours will be needed for the first-year and 150-240 hours for every year after that for the rule, compared to the SRIA’s appraisal of 120 hours for the first year and 18-36 hours for every additional year. ChatGPT, however, puts first year costs both below the SRIA’s estimates (90 hours) and above it (250 hours), while also putting each additional year’s cost between 40 hours and 150 hours, far higher than the official tally.

• Similarly, for the automated decision-making technology (ADMT) provision, both Claude and Grok significantly exceed the government’s 360-hour upper limit for first year costs, for a range of 450-640 hours and 480-730 hours, respectively. On the other hand, ChatGPT puts ADMT costs lower than official estimates for a range of 80-300 hours. Nevertheless, ChatGPT still places a higher cost for each additional year (40-150 hours) compared to official numbers (36-108 hours).

• Overall, California’s estimates appear to be particularly optimistic about ongoing compliance costs, while ChatGPT, Claude and Grok all predicted much higher ongoing costs.

• For BIS reporting requirements, Claude projects 1,140 hours will be needed for first-year compliance, putting it more than three times the official estimate of 333 hours. Consistent with the two other estimates, ChatGPT predicts lower costs (280 hours) than the government’s projection, while Grok’s estimate (380 hours) aligns closely with the official figure. Importantly, all three LLMs consistently project that ongoing compliance will require substantial resources.

While all three LLMs acknowledge that company size, existing legal infrastructure, and rule complexity will significantly impact the hours required, they differ in how they break down the specific tasks. Grok provides the most granular analysis, detailing specific regulatory requirements from the relevant sections of the proposed rules. Claude offers more comprehensive subcategories of work, particularly in training requirements and technical evaluation aspects. ChatGPT takes a broader approach with fewer categories but provides practical context about why the hours are necessary.

To further the analysis, the LLMs cost estimates for RA and ADMT have been converted into discounted regulatory costs using the wage rates from the CCPA section above. The table below charts those costs.

Again, what’s noticeable is that cost projections vary widely. California’s own SRIA estimates for RA compliance puts the cost between $4.4-$7 billion over ten years, while Claude and Grok project costs 10 times higher at $47.3-$74.1 billion. Even the larger SRIA full compliance scenario ($7.6-$12.0 billion) falls far short of the LLM projections. The gap is even more pronounced for ADMT rules, where Grok’s upper estimate of $177.4 billion is more than six times California’s official upper bound of $28.3 billion. Notably, while ChatGPT provides more conservative estimates than the other LLMs, it still projects significantly higher costs because of higher ongoing costs.

All of these estimates are heroic; there’s no doubt about that. But the LLMs provide a gut check, and consistent among them all, it seems that official projections may be overly optimistic. While LLM-generated estimates cannot be considered definitive, they offer a valuable alternative perspective.

Especially for policy analysts, there’s promise with these tools. Instead of relying on analysts’ best guesses, large language models could simulate diverse scenarios that cost estimates are meant to capture, from the small business owner encountering a complex tax form for the first time to the compliance professional at a large corporation managing routine filings. By creating realistic digital twins that reflect varying levels of expertise, available resources, and familiarity with government processes, we could finally generate paperwork burden estimates that capture the true range of human experiences. These AI simulations wouldn’t replace traditional methods entirely, but they could inject much-needed rigor and consistency into a process that has remained stubbornly subjective for decades.

All of the data can be found here. The complete LLM conversations are available in the following links: ChatGPT CCPA compliance chat; Claude CCPA compliance chat; Grok CCPA compliance chat; ChatGPT BIS compliance chat; Claude BIS compliance chat; and Grok BIS compliance chat.

Friends,

It has been over a year since we last chatted and a lot has happened, so I thought I would send you an update on my life and what will become of this space.

Around this time last year, I switched jobs, becoming a Resident Senior Fellow at the American Enterprise Institute. My ambit hasn’t changed, I’m still closely following tech and innovation policy, but I’m spending most of my time working to ensure we have smart AI policy.

Then last March, I started writing the Techne newsletter at The Dispatch. My goal was to cover tech and innovation policy, which included commentary on state AI bills, autonomous cars, Google’s antitrust cases, the $42.45 billion program for broadband expansion, Nvidia and the value of compute, NASA’s canceled VIPER project, and recently, the competitive implications of DeepSeek. But I also covered topics in geoengineering and explained why LA’s plant ecosystem drove the fires.

But I wrapped up that newsletter a couple weeks ago, largely because of what happened in my personal life. Both of my parents died last year. My dad, Eric Rinehart, passed away on June 19th, and then my mother, Penny Rinehart, followed him three months later. Probably the hardest thing I ever had to write were their obituaries. How can you summarize someone’s life in a thousand words or less? Needless to say, settling their estate has become a part-time job.

It was a rough winter without them. Every year during the holidays, I would head back to Central Illinois and spend time with them, helping around the house after working 9 to 5. And in recent years, I would often talk to my dad to get his take on various projects I was working on. He was an economist for the State of Illinois’ economic development agency and a stickler for proper grammar, so he was the first person I would send a draft to.

There's so much of me that's from them, but if there is anything that truly defines their legacy in me, it's their optimism. Both of my parents loved the spring, my father for the chance to restart his gardens and my mother to bring back to life the pond that centered our backyard. Spring is approaching and I too am excited for renewal.

All of this is to say, Exformation is about to become much more active as it will serve as a primary outlet for my work. Sometimes that will mean something short and technical. Other times, it will be a long form piece that just doesn’t work anywhere else. I’m also going to be sending out some of my older Techne work, updated or with additional context.

Thank you for your patience during this quiet period. As I step into this new season, both professionally and personally, I'm excited to grow this space. As always, please don't hesitate to reach out. After all, the best ideas often come from conversation.

Until next time,

🚀 Will

It is a serious graph because it shows the suicide rate for teen boys and girls aged 15-19. So, I pulled all the data again and discovered my mistake. I accidentally pulled a segment of the data for 13-19 year olds instead of simply 15-19 year olds for 1999 to 2020.

It was not my intention to include this broader group in the graph, but as even Stein expressed, “this is an understandable mistake that anyone could make.”

While the change in the data doesn’t fundamentally change my analysis either then or now, it means the chart should look like this:

For transparency's sake, I thought I would share the full recalculation, walk through the framing of trends and cycles we suggested in the first post, and then describe some surprising trends, in my estimation.

Teen suicide is a very serious public health problem. While my mistake did change some aspects of the graph, the big-picture story laid out in the original post is still the same. Teen suicides rose into the 1990s, saw a decline into the 2000s, and rose again in recent years. Understanding what’s driving these decade-long changes is critical to know how to turn things around.

The calculation

As Stein notes in the Substack post,

Rinehart had to make three separate inquiries corresponding to ICD 8 (1968-78), ICD 9 (1979-1998), and ICD 10 classifications.

The International Classification of Diseases (ICD) is the international standard for comparable statistics on causes of mortality and morbidity. ICD-8, ICD-9, and ICD-10 are successive iterations of the standard.

Pulling each one of these segments is how you construct a graph that includes all available suicide data from the 1960s. ICD-8 matches with ICD-9, so you use E950-E959 for both. Under ICD-10, the term previously known as Suicide was renamed Intentional self-harm, and it was reassigned to a new category X60–X84 in addition to Y87.0.

This is the step of the process where I made a mistake. The original graph selected for X60–X84 in 13-19 year olds when it should have been 15-19 year olds in both X60–X84 and Y87.0.

It remains a crude rate and not an age-adjusted rate because I wasn't comparing geographic areas but the entire United States against itself over time. But this could be an incorrect assumption.  

Altogether then, the data was constructed via segments from

• 1968-1978, saved here;

• 1979-1998, saved here; and

• 1999-2020, saved here.

Results were grouped by Age Group, Year, and then Gender. This updated Google Sheet contains the original as well as updated data.

After pulling the data several times, I confirmed that the old graph uses data for 13-19-year-olds instead of 15-19-year-olds from 1999 to 2020.

As Stein rightly point out,

Neither Rinehart nor Barkley are psychologists or public health experts, and the graph was published in a mere blog post, not in a peer-reviewed journal.

Indeed, this is correct! I am a trained economist, and the graph was from a blog post we published that was meant to start exploring this topic. I never intended it to be the absolute word on the subject, but was trying to begin to build a body of work and better understand the dynamics of play. Part of that post also teased a framing of the issue as trends and cycles. More on that later.

The original looks like this:

Instead, It should look like this:

This also comports with the Twenge post here cited by Stein.

Stacking the new and the old data:

What’s changed between the older, incorrect graph, and the updated version:

• The total suicide rate reached an absolute peak in 2017, although as of 2020, it is still below all-time highs. 

• The total rate, boys rate, and girls rate are higher.

What hasn’t changed between the two graphs:

• The early 1990s was a relative highpoint in teen suicides.

• The mid to late 2000s were a relative lowpoint in teen suicides.

• Suicide rates are again rising.

• Girls suicide rates are the highest they have ever been, this was the case with the previous graph as well.

Trends and cycles

The dominant framing in the original post contrasted trends with cycles. Here is what we said,

Is what we are seeing a trend or a cycle?

Trends are different from cycles. Cycles revert towards the mean. They fluctuate around an average point over years. Trends, on the other hand, are something new. They mark a change that is out of the normal cycle.

While I understand there can be confusion over it, the binary framing of a trend versus a cycle highlights one way to think through policy questions. I was attempting to colloquially explain the process of isolating the true effect from everything else. Is what we are seeing a trend, or something else, a cycle?

But I can understand why this framing of trend and cycle is confusing. Indeed, even in that second paragraph, I define cycles as both reverting towards the mean, which is a strictly defined financial term for an asset’s price to converge over time, as well as fluctuating towards an average point, which can suggest a moving average process or even variance.

To add to this, cycles can sometimes be confused with cyclical components which has strict definitions. Sometimes, cycles are even used to describe seasonality. Cycles, as a term, is intensely polysemic.

If I were to rewrite the original piece, I would focus on the binary of trends versus everything else, and just do away with all of the baggage of cycles. Of course, the trend that is driving the conversation is how technology is contributing to mental health changes, holding all else constant.

But not everything is constant.

The school season

One of the more surprising parts of diving into the data was realizing the seasonality in teen suicides. Teen suicides tend to parallel the school year with peaks in April and October, and a significant drop in July. But there have always been deep questions about this relationship because spring peaks are common for adults as well.

COVID seems to have confirmed the association between school and suicide.

COVID acted like a natural experiment. Schools were unexpectedly closed when they should been opened. And the closures happened just when the suicide peak should have occurred but didn’t.

In compiling emergency department visits and hospitalizations for suicidal ideation and suicide attempts, Kim, Krause & Lane (2023) found, “the presence of seasonal patterns and an observed unexpected decrease in suicidality among children and adolescents after COVID-19–related school closures in March 2020.” Their work underscores the association between suicidality and the school calendar, which is displayed in the graph from their report below.

Notwithstanding, the rise in suicide rates is troubling. Grasping the underlying factors behind the decade-long changes is essential for devising effective interventions.

The year is about to close out, and still, a federal privacy law remains out of reach of Congress. Twelve states now have privacy bills, seven of which passed this year.

Earlier this semester, I talked with a student fellow about this stalemate between the states and Congress on privacy, and to help him understand the friction, I gamed out the scenarios. 

There are two scenarios, option A and option B. Option A is the baseline; it's what is happening right now. In option A, privacy bills are passed in each state house, expanding from the current 12 states to all 50 states. Option B involves a negotiated federal privacy bill in Congress.

There is precedent for option A in the development of data breach notification law in the United States. Back in 2002, there were zero data breach laws. These laws force companies to announce when they have been breached. Slowly, one by one, states started adopting data breach notification laws. Now, every state has one.   

Option B is a negotiated agreement in Congress, a federal privacy bill. There have been calls for privacy legislation going back decades now. Still, the fact that Congress hasn’t passed anything underscores how much political capital would have to be spent to achieve an agreement. 

The most recent vehicle for privacy has been the American Data Privacy and Protection Act or ADPPA. (Find the text of the bill here.) A brief history of the bill is helpful to understand the tensions.

Sometime in January 2022, Democrats in Congress reached out to the Republicans on privacy. What came from this discussion was ADPPA, which was jointly released on June 3, 2023, by key members of the House and Senate Commerce Committees. It would establish a comprehensive consumer data privacy framework. 

Democrats had control of both chambers in 2022 with a slim majority in the Senate, and still, they couldn’t get privacy passed. By all accounts, Senator Cantwell is the holdup on privacy.

As explained in Axios, 

The so-called four corners discussion between Pallone, Rogers, Wicker and Senate Commerce Committee Chairwoman Maria Cantwell (D-Wash.) broke down over concerns Cantwell had about enforcement and preemption, according to an aide.

Preemption has always been a sticking point. Under ADPPA's preemption, Illinois could keep its biometric privacy law, but every other state and city statute would be struck down, including California’s privacy bills. Option B would prevent option A. In other words, the two choices create a binary.

So why not just let it ride and allow states to create a patchwork of privacy bills? What’s the harm in option A?

If anything is clear about this space, it is that privacy bills impose costs, which I compiled here.

Given that every bill is slightly different in construction, being unique in breadth and specificity,  compliance will differ from state to state. Each state privacy bill means additional costs for firms that must comply with the law, like Meta and Google. It also means additional litigation risk. Complying with 50 versions of a privacy regime is costly and risky, so the litigation risk simplifies by forcing this onto the federal level. 

I remember talking to one person in a big tech legal team, and they said something I will never forget about the patchwork of state privacy bills.

We will be defensible, but I am not sure we could ever be technically compliant. 

Their quip, however, speaks to a more fundamental feature of privacy bills. Any firm getting caught up in the scope of a bill will need to pursue technical compliance. But, it is an elusive target for even the best-paid and most legal teams.

If police are working on an investigation and want to tap your phone lines, they’ll effectively need to get a warrant. They will also need to get a warrant to search your home, your business, and your mail. 

But if they want to access your email, all they need is just to wait for 180 days.

Because of a 1986 law called the Electronic Communications Privacy Act (ECPA), people using third-party email providers, like Gmail, only get 180 days of warrant protection. It’s an odd quirk of the law that only exists because no one in 1986 could imagine holding onto emails longer than 180 days. There simply wasn’t space for it back then!1

ECPA is a stark illustration of consistent phenomena in government: policy choices, especially technical requirements, have durable and long-lasting effects. There are more mundane examples as well. GPS could be dramatically more accurate but when the optical system was recently upgraded, it was held back by a technical requirement in the Federal Enterprise Architecture Framework (FEAF) of 1999. More accurate headlights have been shown to be better at reducing night crashes yet adaptive headlights only just got approved last year, nearly 16 years after Europe because of technical requirements in FMVSS 108. All it takes is one law or regulation to crystallize an idea into an enduring framework that fails to keep up with developments.

I fear the approach pushed by the White House in their recent executive order (EO) on AI might represent another crystallization moment. ChatGPT has been public for a year, the models on which they are based are only five years old, and yet the administration is already working to set the terms for regulation. 

The “Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence” is sprawling. It spans 13 sections, extends over 100 pages, and lays out nearly 100 deliverables for every major agency. While there are praiseworthy elements to the document, there is also a lot of cause for concern.

Among the biggest changes is the new authority the White House has claimed over newly designated “dual use foundation models.” As the EO defines it, a dual-use foundation model is 

an AI model that is trained on broad data; generally uses self-supervision; contains at least tens of billions of parameters; is applicable across a wide range of contexts; and that exhibits, or could be easily modified to exhibit, high levels of performance at tasks that pose a serious risk to security, national economic security, national public health or safety, or any combination of those matters.

While the designation seems to be common sense, it is new and without provenance. Until last week, no one had talked about dual use foundation models. Rather, the designation does comport with the power the president has over the export of military tech. 

As the EO explains it, the administration is especially interested in those models with the potential to

• lower the barrier of entry for non-experts to design, synthesize, acquire, or use chemical, biological, radiological, or nuclear (CBRN) weapons;

• enable powerful offensive cyber operations through automated vulnerability discovery and exploitation against a wide range of potential targets of cyber attacks; or

• permit the evasion of human control or oversight through means of deception or obfuscation

The White House is justifying its regulation of these models under the Defense Production Act (DPA), a federal law first enacted in 1950 to respond to the Korean War. Modeled after World War II’s War Powers Acts, the DPA was part of a broad civil defense and war mobilization effort that gave the President the power to requisition materials and property, expand government and private defense production capacity, ration consumer goods, and fix wage and price ceilings, among other powers. 

The DPA is reauthorized every five years, which has allowed Congress to expand the set of presidential powers in the DPA. Today, the allowable use of DPA extends far beyond U.S. military preparedness and includes domestic preparedness, response, and recovery from hazards, terrorist attacks, and other national emergencies. The DPA has long been intended to address market failures and slow procurement processes in times of crisis. Now the Biden Administration is using DPA to force companies to open up their AI models. 

The administration's invocation of the Defense Production Act is clearly a strategic maneuver to utilize the maximum extent of its DPA power in service of Biden’s AI policy agenda. The difficult part of this process now sits with the Department of Commerce, which has 90 days to issue regulations.

In turn, the Department will likely use the DPA’s industrial base assessment power to force companies to disclose various aspects of their AI models. Soon enough, dual use foundation models will have to report to the government tests based on guidance developed by the National Institute of Standards and Technology (NIST). But that guidance won’t be available for another 270 days. In other words, Commerce will regulate companies without knowing what they will be beholden to. 

Recent news from the United Kingdom suggests that all of the major players in AI are going to be included in the new regulation. In closing out a two-day summit on AI, British Prime Minister Rishi Sunak announced that eight companies were going to give deeper access to their models in an agreement that had been signed by Australia, Canada, the European Union, France, Germany, Italy, Japan, Korea, Singapore, the U.S. and the U.K. Those eight companies included Amazon Web Services, Anthropic, Google, as well its subsidiary DeepMind, Inflection AI, Meta, Microsoft, Mistral AI, and OpenAI.

Thankfully, the administration isn’t pushing for a pause on AI development, they aren’t denouncing more advanced models, nor are they suggesting that AI needs to be licensed. But this is probably because doing so would face a tough legal challenge. Indeed, it seems little appreciated by the AI community that the demand to report on models is a kind of compelled speech, which has typically triggered First Amendment scrutiny. But the courts have occasionally recognized that compelled commercial speech may actually advance First Amendment interests more than undermine them.

The EO clearly marks a shift in AI regulation because of what will come next. In addition to the countless deliverables, the EO encourages agencies to use their full power to advance rulemaking. 

For example, the EO explains that,

the Federal Trade Commission is encouraged to consider, as it deems appropriate, whether to exercise the Commission’s existing authorities, including its rulemaking authority under the Federal Trade Commission Act, 15 U.S.C. 41 et seq., to ensure fair competition in the AI marketplace and to ensure that consumers and workers are protected from harms that may be enabled by the use of AI.

Innocuous as it may seem, the Federal Trade Commission, as well as all of the other agencies that have been encouraged to use their power by the administration, could come under court scrutiny. In West Virginia v. EPA, the Supreme Court made it more difficult for agencies to expand their power when the court established the major questions doctrine. This new line of legal reasoning takes an ax to agency delegation. Unless there's explicit, clear-cut authority granted by Congress, an agency cannot regulate a major economic or political issue. Agency efforts to push rules on AI could get caught up by the courts.

To be fair, there are a lot of positive actions that this EO advances.2 But details matter, and it will take time for the critical details to emerge.

Meanwhile, we need to be attentive to the creep of power. As Adam Thierer described this catch-22,

While there is nothing wrong with federal agencies being encouraged through the EO to use NIST’s AI Risk Management Framework to help guide sensible AI governance standards, it is crucial to recall that the framework is voluntary and meant to be highly flexible and iterative—not an open-ended mandate for widespread algorithmic regulation. The Biden EO appears to empower agencies to gradually convert that voluntary guidance and other amorphous guidelines into a sort of back-door regulatory regime (a process made easier by the lack of congressional action on AI issues).

In all, the EO is a mixed bag that will take time to shake out. On this, my colleague Neil Chilson is right: some of it is good, some is bad, and some is downright ugly. 

Still, the path we are currently navigating with the Executive Order on AI parallels similar paths in ECPA, GPS, and adaptive lights. It underscores a fundamental truth about legal decisions: even the technical rules we set today will shape the landscape for years, perhaps decades, to come. As we move forward, we must tread carefully, ensuring that our legal frameworks are adaptable and resilient, capable of evolving alongside the very technologies they seek to regulate.

A heartfelt thanks to Caden Rosenbaum for his meaningful input and assistance in crafting this article.

Joshua Browder’s offer was generous. $1 million is what you could make if you took AirPods into the Supreme Court and said exactly what his program asked. But as soon as the offer was extended, it was retracted by Browder, the CEO of DoNotPay. 

DoNotPay made a name for itself by using AI to help people contest parking tickets, get refunds from flights, and cancel free trials. But the company was making the leap into the real world with its legal services. They had got someone to agree to use their service in a municipal traffic court and Browder was hoping to take the idea all the way to the top court. 

The trial was originally scheduled to take place earlier in the year but just days before it was about to go down, DoNotPay pulled out of the trial. It seems that the threats from the California Bar Association that the company was operating without a law license stopped the project. 

As Bobby Allyn of NPR reported,

In a statement, State Bar of California Chief Trial Counsel George Cardona declined to comment on the probe into DoNotPay but said the organization has a duty to investigative possible instances of unauthorized practice of law.

"We regularly let potential violators know that they could face prosecution in civil or criminal court, which is entirely up to law enforcement," Cardona said in a statement.

AI disruption met the hard wall of the law.

From the 1950s to the early 1990s, the Indian economy was under stringent government control and regulation via a system colloquially known as the Licence Raj or the Permit Raj. Businesses in India had to get all sorts of licenses to operate, and they were never particularly easy to obtain.

In much the same vein, today's technological landscape faces its own set of bureaucratic hurdles, reminiscent of the Licence Raj era. Let’s call it the AI License Raj. This contemporary version is woven through our regulatory structures, courts, and laws, persistently binding biological individuals to the law. The AI License Raj silently dictates terms, making it hard for AI to get adopted. 

For example,

• The U.S. District Court for the Eastern District of North Carolina upheld a North Carolina restriction on drone operators, claiming their maps and models amount to illegal “surveying.” 

• Physicians open themselves to liability concerns if they base their diagnosis on AI systems. As one report explained the current regime, “law incentivizes physicians to minimize the potential value of AI.”

• If adopted, California’s AB-96 would give unions the ability to collectively bargain before any local government deploys automated transit vehicles.

• Meanwhile, there is a massive fight over mandates to require two crew members on freight trains. European countries have gone to one person, and Australia is testing completely robotic systems, but the U.S. seems to be poised to make two a requirement.

• Relatedly, a legal challenge against GitHub Copilot and OpenAI Codex has been filed in a US federal court through a class-action lawsuit. The claim includes accusations against GitHub, Microsoft, and OpenAI for violating open-source licenses. 

In other words, the courts and the administrative state will act as a brake on AI adoption because AI systems will face difficulties in trying to get the proper license to operate.

Late last year, Sam Hammond warned of the flood that was coming with generative AI. To his credit, Hammond is right that “AI driven services [will] upend 20th century transaction cost structures and multiply the throughput demanded of our legacy institutions many thousands fold.” 

But I think he takes it a bit too far in worrying that, “The courts, meanwhile, will be flooded with lawsuits because who needs to pay attorney fees when your phone can file an airtight motion for you?” 

Only meatspace attorneys can file motions. An AI version would be immediately thrown out by the court. Indeed, this is exactly why DoNotPay pulled out of the experiment.

If they are to be deployed widely, AI systems will have to disrupt a massive system of licensure industry by industry. 

Still, Hammond is right that institutions will need to change. Indeed, my latest op-ed is all about this. It is about the possibility of using AI to clean up government. 

ChatGPT needs to be turned on the government. A ChatGVT is needed. 

A ChatGVT could take any number of forms, as I wrote,

It could provide straight answers about the newest tax plan, if a bill is stuck in committee, or the likelihood that a piece of legislation will pass. Or a ChatGVT could be turned on the regulatory code to understand its true cost to households and businesses.

Understanding how laws, litigation, hearings, regulatory codes, and administrative actions intermingle can elude even the most experienced experts. The newest generation of Large Language Models (LLMs) appear to be quite effective at working through text with a little bit of tuning.

Using AI to turn law into code will mean that the true impact of government will be understandable  and accessible. Most know that the burden imposed by regulation is colossal but the exact costs are hard to quantify. A ChatGVT could help sort out that problem.

Turning all of that text into computer code will also make reform easier because we will be able to subject it to software management practices, like refactoring. In refactoring, an existing body of computer code is simplified without changing its functional behavior.

A ChatGVT could be focused on refactoring the U.S. regulatory code.

Precedent already exists. During the Trump Administration, the U.S. Department of Health and Human Services (HHS) undertook a program to root out outdated and ineffective laws using AI tools. As a result of this project, HHS cleared a bunch of regulations from the books.

AI tools could also help government agencies upgrade their technology.

COVID made abundantly clear the cost of running on old systems. Old, outdated government systems meant that it became tougher to catch fraudsters as the claims rolled in. While numbers are hard to come by, estimates suggest that unemployment insurance paid $60 billion in fraudulent charges in 2021 alone.

Part of this cost is due to a problem known as technical debt. Businesses accumulate technical debt when they overlook infrastructure issues that could cause future complications. Governments also accrue technical debt when program infrastructure isn’t updated over time.

AI tools could make it much cheaper to upgrade to newer, more secure programming languages to fight waste, fraud, and abuse. Telecommunications and financial services firms have long used computer-aided technologies to detect transactional fraud, money laundering, identity theft, and account takeovers. Governments should be adopting these tools.

In an interview some years ago, the novelist Richard Powers responded with two lines that have always stuck with me. While he was talking about virtual reality, it felt as though he was talking about public policy:

The reason why the public is seduced by virtual reality, why it has embraced this fantasy of the disembodied self, is the desire for the ahistorical, the disembodied will. There is something in us that loves the idea of placing ourselves into immortal space, where our wishes can be met without the drag and impediment of stuff.

Tech policy especially is driven by “the desire for the ahistorical, the disembodied will.” Everyone loves the fantasy, “of placing ourselves into immortal space, where our wishes can be met without the drag and impediment of stuff.” 

But the real world is full of impediments. It is full of stuff. When it comes to AI, this is where our attention should be focused. 

Of course, you need to focus on power, but also remember what Kenneth Burke said: A way of seeing is also a way of not seeing—a focus upon object A involves a neglect of object B.

Power is one frame for understanding, but it is one model among many. And to be clear, the term model does a lot of work. It is encompassing and includes a range of related ideas. Models, frames, frameworks, shorthands, or heuristics are all tools and organizational schemas that help to break down and simplify complex data.

Those of us working in public policy clearly traffic in models. Rather than being coy about it, we must embrace our models, interrogate them, and understand their limits. 

Two heuristics that I cannot shake stem from the work of Friedrich Hayek and James Buchanan. The first concerns information, and the second concerns process. Both are important views to adopt when understanding the opaque world of politics and policy.

Hayekian welfare states

Friedrich Hayek's research aimed to understand how knowledge gets embedded into economic systems. His key insight, which netted him a Nobel, was that knowledge is naturally dispersed throughout society.

Reading Andreas Bergh (2019) last year brought this into focus for me. This paper asks two questions. First, if big government harms economic and social outcomes, why aren't the nations with the most intrusive governments doing worse? And following this, why don't more countries follow the strategy of Sweden and other Scandinavian states to provide for their citizens?

To answer those questions, Bergh introduces the concept of Hayekian welfare states and argues that it can help to answer both. A Hayekian welfare state refers to the “tendency to evolve through trial and error and the tendency to have a large public sector without being highly exposed to the Hayekian knowledge problem described by Hayek (1945).”

Politics and policy have long been dominated by a big versus small government divide. At the end of the day, this boils down to fiscal capacity. In this thinking, we will tend to ask, how much does the government take from your wallet? 

A Hayekian understanding of government is orthogonal to the fiscal question and focuses our attention on knowledge within the system. It is a separate line of analysis from the fiscal one. Bergh makes it clearer with this table which puts the fiscal capacity question vertically and the Hayekian knowledge question vertically: 

Few appreciate how little information sharing there exists in many government institutions, but Matthew Yglesias had a great example of it:

I remember showing this chart about Democrats’ paid leave proposal to a friend who works in political journalism. She was genuinely very surprised — she thought the proposal was for universal parental leave and had no idea that over 60 percent of the benefits were for personal sick leave or that 30 percent of new mothers wouldn’t qualify for coverage.

After that, I surveyed some members of Congress and chiefs of staff I know on the Hill, and about half of them didn’t know either. This was not secret information — it was in the CBO score — but the fact that Democrats had so little information about the content of their own policy was a sign, I think, of how impoverished the policy debate has become. Nobody was out there making the case that “hey, the part of this that people are fired-up about is the leave for new parents, let’s narrow the bill but make the coverage universal and it’ll be cheaper.” And because nobody was making that case, nobody was making the affirmative case for the structure they decided on.

Indeed, even the most well-informed are uncertain of outcomes. As Jason Furman explained in Maxims for Thinking Analytically, when he arrived in Washington, he was surprised to learn that,

The best political and legislative aides had only two possible probabilities for events occurring: zero and one. One of the top ones would say ‘I was in Congress for years, and I can tell you there is no chance they will pass this’ – weeks before it passed. Or ‘In all my decades in Washington attaching such and such to the bill has never failed to get it passed,’ just before months and months of failure to get it passed.

Bergh (2019) fits squarely within a literature first laid down by Herbert Simon. Indeed, as I have noted before, my view of information is indelibly shaped by Simon’s “Designing organizations for an information-rich world.” His logic is generative because it games out the impacts of an information-rich world. And it is driven by a simple yet powerful question: In a surfeit of information, what gets economized?

Simon’s reasoning is always worth quoting in full:

[I]n an information-rich world, the wealth of information means a dearth of something else: a scarcity of whatever it is that information consumes. What information consumes is rather obvious: it consumes the attention of its recipients. Hence a wealth of information creates a poverty of attention and a need to allocate that attention efficiently among the overabundance of information sources that might consume it.

In applying this logic to individuals, Simon was the first to name and describe the attention economy. And in applying this logic to organizations, Simon brought attention to the processes that condense information for groups and organizations. The key question for understanding information in systems is to understand what is being “withheld from the attention of other parts of the system.”

This is the driving question of our age. What is being “withheld from the attention of other parts of the system” is a throughline behind all of our contemporary scandals like Cambridge Analytica, Hunter Biden’s laptop, Section 230 debates, shadowbanning, content moderation, free speech online, censorship by Big Tech, and on and on.

But thinking about how information is being withheld is also a powerful lens to understand how policy and politics work in practice.

Policy is the result of a competitive process.

The other powerful heuristic I saw everywhere this past year comes from James Buchanan. His classic text on order complements Hayekianism well. "Order Defined in the Process of its Emergence" is a short piece but an incredibly important turn in logic. As he pointed out, the order of the market “emerges only from the process of voluntary exchange among the participating individuals."

Buchanan continued that “the allocation-distribution result, does not, and cannot, exist independently of the trading process. Absent this process, there is and can be no ‘order.'”

All of this clicked for me when I read Marta Podemska-Mikluch's paper on Polish pricing of reimbursed pharmaceuticals which reframes Buchanan for public policy:

By assuming that policy is an object of choice, economists have no alternative but to naively hope for a decision-maker sensitive to economic logic. An alternative approach is to think of policy, not as an object of choice but as an outcome of a competitive process.

I think that understanding is key to the movement in Congress. Policy is not an object of choice but is an outcome of a political trading process.

It is a bit of a banal observation to say that policy is an outcome of a competitive political process, but the shift in framing helps to clear up some confusion. As Podemska-Mikluch continues,

In some cases, it is a benevolent dictator, in others a special interest group, but most commonly a median voter is thought to be responsible for choosing a policy from options proposed by various individuals and groups within the society. No matter which kind of a decision maker is considered, the implication is the same: policy is an object of choice. Conceptualized in this manner, the process of policy formation is not much different from shopping at a supermarket where one picks and chooses the ingredients for the optimal policy (Wagner 2010). As a result, much of the story is missed, i.e. unintended consequences that are generated in interactions between individuals pursuing their plans and goals.

If policy were a choice made by the median voter, we’d have immigration reform, and yet, not even the EAGLE Act can get past Congress. Immigration is more popular than ever, even as its passage seems less likely. What’s stopping immigration reform are all the veto players in the process. 

Legislation is the end product of the committee process, at least in Regular Order. Each time a bill can pass a checkpoint, it gains steam and becomes more likely. Passing out of committee and onto the House floor, for example, means that the bill has a better chance of becoming law. The more steps you complete on the board, the more likely the bill becomes. 

So creating legislation isn’t as simple as deciding the correct choice. Instead, it is an emergent phenomenon produced from the competition of various actors, all jostling for status, constituent support, legal clarification, and prohibitions within a competitive institution. And then these bills have to get interpreted through the administrative state.

Seeing policy as the outcome of a competitive process explains the power of Senator Manchin. In the last session, he had a coveted position of power because he was the most conservative member in the Senate in a 50-50 split with the Republicans. But his dealings on his permitting bill show the limits of that veto power. As Politico reported,

Manchin’s permitting bill began as a coda to his outsized leverage in a Congress that found him playing decisive roles in everything from a bipartisan infrastructure law to post-Jan. 6 presidential certification reform to two massive Democratic-only bills. The centrist hatched a two-part deal with Senate Majority Leader Chuck Schumer this summer: First Manchin would help pass a party-line climate, health care and tax bill, then Schumer would take up a plan to expedite big energy projects including West Virginia’s own Mountain Valley Pipeline.

But Manchin’s final major priority after a stretch in which everything broke his way needed the support of Republicans. And there were simply too many problems for him to solve in too short a time after releasing his legislation just last week. His home-state GOP colleague Sen. Shelley Moore Capito has her own permitting bill, and Republicans who want to defeat Manchin in 2024 largely have no desire to help him out of a jam.

“He thought he was going to pass a bill and get it signed into law. He miscalculated, is the nicest way I could put it,” said Sen. John Cornyn (R-Texas), a close ally of Senate Minority Leader Mitch McConnell, who whipped against Manchin’s effort behind the scenes and publicly pushed for its defeat on Tuesday.

If we move away from a pure power lens and towards one considerate of knowledge problems and production problems, then Manchin’s problems in permitting reform make sense. Manchin was the final vote only in a negative sense. He didn’t have the clout to advance all of his bills, but he did have the ability to take provisions away. As Jim Newell explained in Slate, “He’s kind of being squeezed because he is not determining the agenda, but he is determining what may or may not cut through as part of that agenda.” 

Determining what may or may not cut through is an underappreciated part of government operations. This is why I spent so much time working on vetocracy last year because in many political processes, “too many actors have veto rights over what gets built.”

As I wrote,

Vetocracy is related but separate from the broader problem of red tape. Red tape, permitting, and other limitations regulate conduct. Laws that regulate conduct need to be judged by their own merits, but vetocracy isn’t regulation of conduct. Vetocracy is about the needless delays created through excessive veto points throughout our institutions. There are a lot of players involved in permitting and all it takes is one to slow everything down. Vetocracy is about the excessive accretion of voice that slows down normal processes.

In other words, public policy is not as much a problem to be solved as it is the result of a process. And that process can be slowed down by veto players. Going slow is costly.

In a recent research roundup, I highlighted Neupane and Adhikari's (2022) study on permitting costs. These two researchers modeled the costs of 11 geothermal projects in California, Nevada, and Utah and then calculated how the variation in review times can change the viability of a project. Longer CEQA/NEPA review timelines mean less time to earn returns, pushing the cost upwards. The effect was an electricity cost increase by 4% to 11% via the simplified levelized cost of electricity (sLCOE) method. These extended review timelines led to $64 million to $227 million in revenue losses. For some projects, these extended delays meant the death knell. Slow permitting could sink a geothermal project.

Just as permitting processes can affect the viability of a geothermal project, understanding the interplay between politics and policy often requires digging beneath the surface to uncover the underlying processes.

These two heuristics aren’t the only ways to understand government processes, nor are they revelatory, but they are underappreciated. Adopting multiple perspectives, including those centered on information and process, is the only way to gain a more comprehensive and nuanced understanding of the opaque world of politics and policy.

Share

1. An electric scooter ban increased congestion by 10% in Atlanta.

2. Plants are green because it reduces noise in photosynthesis. They ignore the most energy-rich part of sunlight because stability matters more than efficiency.

   

3. Neupane and Adhikari's (2022) study is a must-read on permitting costs. These two researchers modeled the costs of 11 geothermal projects in California, Nevada, and Utah and then calculated the compliance impacts. Longer CEQA/NEPA review timelines result in a 4% to 11% increase in the simplified levelized cost of electricity (sLCOE). Extended review timelines led to revenue losses of $64 million to $227 million. In other words, slow permitting could sink a geothermal project.

4. The tomato variety found in U.S. grocery stores was selected because the variety was hard and could be picked by machine.

5. Sierra Mixe is a rare variety of maize that has evolved a way to make its nitrogen, which could revolutionize farming. I cannot describe how cool I find this gooey corn. Extra: The Wiki entry on fertilizer is stacked.

   

6. Yikes: "Residential segregation alone explains more than 100 percent of school segregation in the U.S.” Tomás E. Monarrez (2022)

7. Why are some stories more successful? Semantic progression might explain it: “Movies and TV shows that move faster are liked more, TV shows that cover more ground are liked less. Academic papers that move faster are cited less, and papers that cover more ground or are more circuitous are cited more.”

8. Research finds that searches for school bullying and cyberbullying are good proxies for the real behavior, and they dropped 30-35 percent as schools shifted to remote learning in spring 2020.

9. Preference cascades might explain why public opinion changes so rapidly. Glenn Reynolds: “Average people behave the way they think they ought to, even though that behavior might not reflect their own personal feelings. Given a sufficient ‘A-HA!’ moment when they discover that their personal feelings are shared by a large portion of the population their behavior may change dramatically.”

10. Consciousness exists at the knife edge (of near criticality): "We show that the electric activity of the cortex is indeed poised near the boundary between stability and chaos during conscious states and transitions away from this boundary during unconsciousness and that this transition disrupts cortical information processing."

11. The U.S. economics profession constantly overstated the growth rates of the Soviet Union. In their best years, Soviet growth was 4%, much below the predicted rates. See Nintil - The Soviet Union: GDP growth and Soviet Economic Growth: 1928-1985

    

    
    

12. NASA’s plans after Apollo: develop a lunar orbit station in 1978, plant a lunar surface base in 1980, and go to Mars in 1981 or 1983. Extra: My ongoing list of space and defense projects.

13. Do heart attacks make you eat avocados? I like the way that Josh Dean flipped the correlation/causation idea, what he calls Correlational/Causal claim Turing test: “If you're not willing to flip the order of your ‘association’ (e.g. Heart Attacks predict consuming avocados), you're making an implicitly casual claim and should either knock it off or be up front about it.”

14. Geniuses of the past were aristocratically tutored. Could AI tutors bring back an age of geniuses? 

15. Is behavioral economics having a reckoning? PNAS: “Our results show that choice architecture interventions overall promote behavior change with a small to medium effect size of Cohen’s d = 0.43 (95% CI [0.38, 0.48]). In addition, we find that the effectiveness of choice architecture interventions varies significantly as a function of technique and domain.”

16. Dare Obansanjo: "The 'Cash me ousside, how bout dah? girl from the viral Dr. Phil video from a few years ago made $52M from OnlyFans in one year ($42M after fees). This is another business category I simply wouldn’t have believed possible if someone told me about it five years ago." I am fascinated by the new creator/attention economy.

17. “While overall income inequality rose over the last five decades, the rise in overall consumption inequality is small.”

18. Top income taxation reduces everyone’s income, not just those at the top. If we bake in some simple assumptions about innovation, then increasing top income taxation reduces everyone’s income, not just income at the top. Enomist Chad Jones’ newish paper relies on three assumptions: (i) new ideas drive economic growth, (ii) the reward for successful innovation is a top income, and (iii) innovation cannot be perfectly targeted by a research subsidy. Combined, these three forces sharply constrain both revenue-maximizing and welfare-maximizing top tax rates.

19. Johnny Appleseed mostly spread seeds for apples used in making cider.

20. Americans now work 50 percent more than do the Germans, French, and Italians. But that wasn’t true fifty years ago. Edward Prescott (2004): "The surprising finding is that [the marginal tax rate’s effect on labor income] accounts for the predominance of the differences at points in time and the large change in relative labor supply over time with the exception of the Italian labor supply in the early 1970s."

21. Metal corrosion costs 3% of global GDP, according to corrosion engineers (2016). Big if true.

22. People dramatically undervalue the Covid-19 vaccine. NBER: “The willingness to pay for initial vaccination is around $50, only 2% of the WTP implied by standard VSL calculations…While standard economic models imply that vaccines are undervalued because of their large externalities, we interpret the finding that WTP estimates are well below the VSL benchmarks as evidence that internalities play a substantial role.”

23. A poll from New York Times and Siena College found that voters choose climate change as their top policy priority around 1 percent of the time. “Just 1% of voters named climate change as the most important issue facing the country, far behind worries about inflation and the economy. Even among voters under 30, the group thought to be most energized by the issue, that figure was 3%.”

24. The widespread cultural diffusion of knowledge started 400,000 years ago. Different groups of hominins probably learned from one another much earlier than was previously thought, and that knowledge was also distributed much further. "To date it was always thought that cultural diffusion actually started only 70,000 years ago when modern humans, Homo sapiens, started to disperse. But the record for the use of fire now seems to show that this happened much earlier," archaeologist and researcher Katharine MacDonald explains.

25. Are commercially oriented societies more virtuous? It seems so. In 543 experiments across 13 villages in Greenland, more market participation meant more moral behavior towards others and more universalism in moral decision-making. 

26. You know that look of Art Deco steel, which has a dull finish? That is called Monel. The Nickel Institute has this interesting piece, "Historic Monel: the alloy that time forgot." The Wiki is good as well.

    

27. The cube root rule is an observation in political science that the number of members in the lower house tends to be the cube root of the population being represented. Under this rule, the U.S. House should have around 637 members. 

28. Radiologists: “Variation in skill can explain 39% of the variation in diagnostic decisions, and policies that improve skill perform better than uniform decision guidelines.”

29. Military innovations, like cavalry and iron weapons, enabled societies to over-power rivals and build strong, bureaucratized states, according to new research. 

30. Why batteries are horrible to eat: "The accumulation of hydroxide ions rapidly increases the surrounding tissue environment to a local tissue pH of 12 to 13. This highly alkaline environment then creates an ensuing liquefactive necrosis..."

31. Expanding the size of the police force saves Black lives, reduces the number of arrests for serious crimes, and reduces the number of serious crimes. But, it increases arrests for petty crimes.

32. Same data, different conclusions: Seventy-three independent research teams used identical cross-country survey data to test an established social science hypothesis: that more immigration will reduce public support for government provision of social policies. But instead of convergence, teams’ numerical results varied greatly, ranging from large negative to large positive effects of immigration on public support. Similar results were found in studies of gender and professional status on verbosity during group meetings, whether soccer referees are more likely to give red cards to dark-skin-toned players than to light-skin-toned players, and how the same MRIs were analyzed by different teams.

33. The Equality Paradox: richer kids are sadder kids. Analysis of 2018 PISA data from nearly half a million 15-year-olds across 72 countries indicates a negative log-linear relationship between per-capita GDP and adolescent life satisfaction. The effects were more pronounced for girls than for boys. Another paper using this same dataset found that gender equality enhances boys’ but not girls’ subjective well-being. It seems that greater gender equality may facilitate social comparisons across genders.

34. Eye contact marks the rise and fall of shared attention in conversation: “Eye contact may be a key mechanism for enabling the coordination of shared and independent modes of thought, allowing conversation to both cohere and evolve.”

35. Amazon Alexa is a “colossal failure.” The division lost $10B in 2022.

36. Daisugi is an ancient Japanese forestry technique in which planted cedars are pruned in a special way to produce "shoots" that eventually become perfect, straight, knot-free lumber.

37. The U.S. highway system was partly advocated because policymakers worried about nuclear attacks, according to the official history of the FHWA.

38. Pants are a combinatorial innovation: “Inventing pants was a matter of combining different weaving techniques from cultures thousands of kilometers apart.”

39. Apparently, “Most of the association between age and social conservatism is accounted for by parenthood.” 

40. Being agnostic is different than being atheist: “Agnostics were more neurotic, but also more prosocially oriented and spiritual, and less dogmatic.” They “compose a distinct psychological category and are not just closet atheists.”

41. The past is a foreign place. "The East India Company traded about 50 tons of tea a year at the start of the nineteenth century and 15,000 toward the end of it." Large ships today can carry ~25,000 tons, on their own. From The End of the World Is Just the Beginning.

42. Children and young adults aged 7 to 29 can multitask. Adults cannot. In older participants, there is a cost for multitasking.

43. D-Day was all about logistics. Only 14% of the D-Day force were infantry, and 6% were tank crews. The rest were "rear area functions,” including logistics, medical staff, and all the functions to make war possible, according to Adam Tooze.

44. Public lawns get about the same weekly hours of use as tennis courts. The top-use spaces are gyms (688 hours), pools (301 hours), and walking loops (345 hours). 

    

45. Over two-thirds (67.4%) of US regulations have never been updated since they were first passed. We need to refactor our regulatory codebase. 

46. Deep Blue likely won against Kasparov because it chose a random move deep in gameplay. Kasparov interpreted the move as some artful tactic and lost his momentum. 

    

47. Economist David Autor and coauthors found that Paycheck Protection Program loan guarantee was expensive at $170-257K/job-year retained, regressive in that 3/4 of funds went to the top 5th of households, and badly targeted since 25-34% went to workers who would've otherwise lost jobs.

48. A lot of places around the US require buildings to have two staircases, even for small apartment buildings.

49. Yo-Der Song and Tomaso Aste found in their paper “The Cost of Bitcoin Mining Has Never Really Increased” that, “Despite a 10-billion-fold increase in hashing activity and a 10-million-fold increase in total energy consumption, we find the cost relative to the volume of transactions has not increased nor decreased since 2010.” More evidence that the cost of energy constrains Bitcoin.

50. The Great Orme copper mine in Whales was so productive that by 1600BC, there were no other copper mines left open in Britain.

A great example of this comes from King Tutankhamun.

When Carter unwrapped the Egyptian pharaoh, he found two daggers, sheathed in gold, in his wrappings. One of the daggers was made of gold, but that wasn’t the one that sparked interest.

The dagger Carter found on Tutankhamun’s right thigh brought the most attention. It was an object out of time. It was made of iron, but came from the Bronze Age, before iron smelting was common.

Now we know the dagger was made from a meteorite. It was a status symbol. And it was probably valued greater than the gold dagger.

The two daggers can be seen above. On the top is the gold dagger. On the bottom is the iron dagger.

The finely crafted blade of iron is homogeneous and non-rusted. Its handle, fashioned from fine gold, features exquisite cloisonné and granulation work and culminates in a rock crystal pommel. The gold sheath is adorned with a floral lily motif on one side and a feather pattern on the other, concluding with the head of a jackal.

The iron blade was special.

Since a kid, I’ve had a fascination with Tut. Yet, it was during Covid that I began reading more comprehensively about Egypt and its place in the ancient world.

Tut was a relatively minor pharaoh who lived at the end of the Eighteenth Dynasty. This dynasty lasted between 1550 BC and 1292 BC and marked the peak of Ancient Egypt’s power.

People typically think that the height of Ancient Egypt came with the pyramids. But by the time of Tut, the pyramids were already 1000 years old.

Instead, the greatest extent of ancient Egypt existed between 1550 BC and 1200 BC. This is when they became a Great Power. They adopted the new technology of chariots and went on a military campaign to bring the entire region under their control.

By the time of Tut, the Egyptian Empire wrapped along the coast of the Mediterranean and reached as north as it would ever go. Northern Canaan, the land of Retjenu, much of Syria and Mittani, in modern day Turkey, were under the influence of Thebes.

Tut’s world was one of excessive opulence for the time, but it was also a world undergoing great change.

Tut was originally born as Tutankhaten, meaning "Living Image of Aten." But his name was changed to Tutankhamun, "Living Image of Amun," later in life when Egypt moved away from Atenism and back towards a more traditional version of the religion.

Tutankhamun’s (likely) father, Akhenaten, introduced Atenism during his reign. In this variant of the ancient religion, worship was centered around Aten, who replaced Amun as the preeminent solar deity.

To go along with his new worship, Akhenaten built a completely new palace and administrative complex at Amarna Aten.

But as soon as Akhenaten died, Egypt did away with Atenism.

Akhenaten’s name was carved out, and Amarna Aten was abandoned. Only recently have archeologists uncovered these massive buildings.

Tut died young, around 18 or 19, as the leader of a sprawling empire already claiming 1,000 years of history. To our benefit, Tut seems to have been forgotten, buried under Ramesses IV. His tomb remained intact and untouched by robbers until Carter opened it in 1922.

And when Carter opened it, he found the dagger and countless other treasures stacked to the ceiling, including many other small iron objects.

The iron dagger found in Tut’s wrappings was, for a long while, a mystery because iron came late to Egypt. But it is generally thought that Tushratta, King of Mitanni, sent several precious iron objects to Amenhotep III, the grandfather of Tutankhamun.

Tut was likely trying to connect his power as descending from Amenhotep, not Akhenaten. It also doesn’t hurt that iron came from the heavens.

Tut’s blade is truly a rare piece because the total catalog of Bronze Age irons is so limited. Nearly every piece of Bronze Age Iron is made from meteorite. Axes from Syria and China dating back to about 1400 BC, a Syrian pendant from 2300 BC, a Turkish dagger from 2500 BC, and beads from Gerzeh, Egypt from 3200 BC all have levels of nickel and cobalt that line up with meteorites. Other than this blade, the Ugarit axe from Syria is probably the most well-known Bronze Age iron.

Axes have been found far and wide in Arabia, suggesting deep trade patterns. But the hematite axes were traded within a small region. As Vincent Charpentier (2019) explained, “In the heart of the Neolithic Middle East, this innovation was specific to Arabian shores between the Musandam and Qatar peninsulas." What is interesting is that, "Within the Middle East, south-eastern Arabia during the Neolithic engaged in a very original means of production of metal objects, as the latter did not focus on copper, a very malleable and much more available material, but on haematite, which was much harder."

The significance of minerals in the study of history and political power cannot be overstated. Minerals have shaped human societies and political power in countless ways, from the use of copper and iron in ancient times to oil, rare earths, and silica of today.

Iron has a deep history with humanity, just like tin.

I think it is lost on most, the importance of tin, but tin is the key ingredient for bronze. By adding tin or arsenic to copper, metalworkers discovered a way to make the material much harder and easier to cast, creating bronze. This development was a turning point for metalworking techniques and marked the transition into the Bronze Age around 3000 BC. In antiquity, tin was a key commodity.

But tin is also a relatively rare element and was mined only in a select number of areas in the classical world. One of the most important deposits in Spain gave rise to the culture we now call Tartessos.

If you have never heard of Tartessos, read the Wikipedia entry on them. Highly developed both politically and culturally, Tartessos was the first organized state in the Iberian Peninsula. They made absolutely beautiful gold objects like the ones below. Not only is Tartessos believed to be the inspiration for the myth of Atlantis, it is also probable that this civilization was the Tarsish of the Bible.

Just last month, the reliefs of five human faces were found at the ancient Tartessian site of Casas del Turuñuelo and it revealed a lot about this society. We don’t know much about Tartessos because the civilization seems to have vanished seemingly overnight perhaps due to a massive fire. But the picture that has emerged with recent excavations suggests that Tartessos equaled Greece and the Etruscans in art.

All of this is due to tin.

This year especially, I am on a journey to understand better the history of minerals, materials, and energy production. Here are some other things I’ve come across:

• The U.S. Geological Survey has a lot of good primers, including this one on rare earth elements (REEs).

• Scientists have recreated tetrataenite in a lab. If synthetic tetrataenite works in industrial applications, it could make green energy technologies significantly cheaper and ween the US off Chinese rare earths.

• The Bushveld of South Africa is where most platinum-group metals (PGMs) are mined. These six metals include ruthenium, rhodium, palladium, osmium, iridium, and platinum. Zeihan is right, the Bushveld "leaks chromium, iron ore, tin, and vanadium," all of which exist in an unadulterated state. Everywhere else it is found, it exists around copper and nickel. The next biggest area where this place is mined is Norilsk, a Soviet-built Arctic penal colony whose workers toil a mile underground.

• You know that look of Art Deco metal, the kind which has a dull finish? That is called Monel. The Nickel Institute has this interesting piece, "Historic Monel: the alloy that time forgot." The Wiki is good as well.

• Anyone got a good history of the International Nickel Company (INCO)? They seem fascinating, and they made Inconel as well: "Inconel alloys are oxidation-corrosion-resistant materials well suited for service in extreme environments subjected to pressure and heat. When heated, Inconel forms a thick, stable, passivating oxide layer protecting the surface from further attack. Inconel retains strength over a wide temperature range, attractive for high-temperature applications where aluminum and steel would succumb to creep as a result of thermally-induced crystal vacancies. Inconel's high-temperature strength is developed by solid solution strengthening or precipitation hardening, depending on the alloy."

• The Meehanite process was developed in the late 1920s and early 1930s, by the Ross Meehan foundry in Chattanooga, Tennessee. This initial discovery was based on the use of calcium silicide to inoculate irons melted in a controlled manner. This resulted in the development of cast irons of greater strength suitable for critical engineering applications. The Wiki.

• Aluminum comes from two sources, bauxite or clays. Bauxite is energy intensive in its own right, but clays are even more so. According to the USGS, "The United States and most other major aluminum-producing countries have essentially inexhaustible subeconomic resources of aluminum in materials other than bauxite."

• I am also reading up on hydrogen tech and new methods of generating hydrogen more efficiently from water.

• Miscellaneous: "The rise and rise of corrugated iron" from The Economist | Iron as ornament, its origins & Evolution | Mineralogy database | Historical prices from 90 minerals from the USGS ; Wiki list of industrial processes.

Earlier in the morning, I felt a scratchy throat, so naturally, I went searching for a test. After calling countless doctor's offices and drug stores and coming up empty, we finally found a testing center an hour away that could take us. But we would have to wait in line.

After standing out in the cold for hours, we each got two tests, a rapid test and a PCR. The rapid test came back negative in 15 minutes or so. As for the PCR tests, we never got them back.  

The entire process was a complete mess.

And we weren't alone. All across the United States, tests were hard to find. But if you cracked open a newspaper, you could read stories of the Europeans having no problem finding them. The COVID wave hit them just as hard, but they seemed to weather it much better. Germany, in particular, was a standout.

Like many of my other niche fascinations, I became obsessed: Why were rapid tests so hard to find a little over a year ago?

In between projects last year, I worked on this question. And a couple of weeks back, my findings were published over at the Center for Growth and Opportunity at Utah State University.

Share The Exformation Newsletter

This Twitter thread has all the important finds from the paper.

Like others that have looked into the shortages, I found that the Food and Drug Administration (FDA) bears much of the blame for limited rapid tests. But I was surprised that the same story had been repeated every decade since the early 1970s.

Back in 1971, Faraday Labs’ Ova II pregnancy test hit the market, becoming the first direct-to-consumer pregnancy test. But the test wasn’t available for long because, in December 1972, the FDA recalled it, saying that they were “inaccurate, unreliable and prone to give false results.”

Funny enough, the agency still allowed the same tests to be processed in labs. An op-ed by one FDA official was telling because he said that women “do not in general have sufficient training to detect malfunctions.” So in effect, the agency claimed that the kit was only inaccurate because it was in the hands of “laywomen,” who were prone to misreading the results. Later research would confirm that error rates were about the same for labs as for the women taking the tests.

Fast forward a half-century and the same demand for accuracy persisted for rapid tests. The US was the only developed country that required an additional set of trials that proved rapid tests were “usable.” Separate from the clinical study establishing the accuracy of the tests, the FDA required that tests undergo a usability study to prove that the general public can use them.

Usability studies and a bevy of other requirements added serious costs to US rapid tests. The real proof can be found in the prices between us and Germany, which are laid out below. German tests were cheap.

There’s so much more in the paper. It explores the economic incentives behind delays and the legal framework of diagnostic testing regulation. It also does a much better job of explaining why the US and Germany are different. Ultimately, I think the paper makes a strong case for reform.

So what can be done? There are several possible reforms that Congress will need to take the lead. They should:

• Amend the Emergency Use Authorization (EUA) process to consider the cost of time;

• Create clear reporting requirements on EUA approvals;

• Rethink the need for diagnostic usability trials;

• Refocus the standards for clinical trials for diagnostics;

• Formalize rules for diagnostics; and

• Mandate that the agency produce an official report on the costs of decision delays during COVID. 

The FDA needs reform because there is clear evidence of regulatory bias. As Joseph A. DiMasi, Christopher-Paul Milne, and Alex Tabarrok first outlined, FDA divisions have widely varying approval times for drug approvals. Oncology drugs and antivirals get approved much faster than other drugs in neurology and cardiovascular. Truth be told, when I first saw them, the charts from the paper shocked me.

During emergencies, the FDA ought to expedite its processes without compromising safety in bringing diagnostics to market. Although the modifications I’ve laid out in the paper may not entirely resolve indecision, they could significantly contribute towards ensuring that future diagnostics are made available to individuals at the most crucial moments.

But testing isn’t the only place where the FDA needs reform. The agency needs to rethink its method of approving drugs. In upcoming work, I hope to explore what’s happening, what’s at stake, and how to change up the process.

As always, thanks for reading!

The paper can be found here, and a Twitter thread can be found here. Retweets are always appreciated.

Subscribe now