Economist Bryan Caplan conceived of the ideological Turing test, popularized by Tyler Cowen, as a way to conceptually test one’s ability to understand the views of those who are opposed to your own. It’s a form of “steelmanning” - strengthening the robustness of your arguments by pitting them against the strongest version of the counterargument.
Here I attempt this for the view that AI policy should be limited to little or no regulation or restrictions. This isn’t the full fat version, which as Bryan originally structured it, involves a sort of double-blind testing.
One way to evaluate this test is to consider the claims below on their own and determine whether they sound like a credible advocate for a moratorium on AI laws.
[My actual priors included at the very end]
Claim #1
Technological disruption has, in the aggregate and over the long run, consistently produced far greater human welfare gains than losses – and there’s no clear reason to expect AI to be the exception.
Every technology throughout history has involved some disruption. These disruptions can upset prior economic equilibriums, for example, but eventually result in (often remarkable) increases in human flourishing and standards of living.
The industrial revolution was met with resistance from some quarters facing the risk of losing their economic rents (particular occupations, perhaps political coalitions built around a specific economic ordering). Some of this resistance is understandable being that it stems from uncertain change to a known economic and social arrangement. Some jobs are inevitably lost; some patterns of behavior and lifestyles are overturned.
But we need to think in the aggregate and long-term. Every example of a productive technological innovation - unproductive ones eventually disappear without entrenchment - leads to increases in overall human welfare, sometimes incredible increases. Mechanizarion led to a collapse in value of some skills. But broadly led to better fed, longer-lived, more creative people through increasing productivity and flowering of specialization.
AI looks like a tsunami of change breaching our shores, but the history of proliferation of new ideas is overwhelmingly positive and we should (as a moral or utilitarian or democratic matter) want the most people to do the most well.1
Claim #2
Regulating AI now means legislating under a high degree of uncertainty about a still immature technology, risking that today’s incomplete understanding gets locked in as tomorrow’s obstinate steady state.
New policy tailored specifically toward addressing perceived risks of AI would be premature: we’re legislating in the dark. There is so much uncertainty about breakthroughs around the corner; we don’t know the implications. Attempting to structure policy schema based on what we know right now is effectively backwards-looking – under the best conditions it will wrap around yesterday’s understanding and choke off useful discovery, foreclosing maximal social benefits.
While we observe leaps in AI capacity and uses all the time right now, we are functionally at the adolescent stage of this technology. It would be as if 1970s biologists, confronting the then-totally novel recombinant DNA technique, had been locked into restrictive practices based on the most pessimistic risk assumptions – before anyone understood that those rules were unnecessary. Instead, the field governed itself first through evolving voluntary guidelines, with formal regulation following only once more was understood. Within a few years, that same technology produced synthetic human insulin, ending the reliance on animal-derived insulin and increasing access for millions of diabetics.2
Claim #3
Frontier model development draws on the same kind of openly available, generally applicable math and science that the US has always treated as basic research – and basic research is not something we regulate.
Regulating AI development at its core would effectively regulate fundamental scientific research. Discovery around models is a kind of basic research. Not basic in terms of simplicity, but as a term of art referring to the pursuit of intellectual curiosity which leads to the expansion of general knowledge. In US policy, we support this mode of research through favorable tax treatment and public funding. The reason is this kind of core scientific inquiry doesn’t always have obvious commercial applications but in the aggregate benefits everyone by expanding our base of knowledge – sometimes in fundamental ways.3
We regulate pharmaceuticals, but the process of developing specific medicines is built on basic research in biology and chemistry. Understanding protein structure, behavior of genes and genetic mapping, cell signaling pathways, all feed into eventual treatments for specific pathologies. But we do not regulate the former because it is generally applicable knowledge creation that benefits everyone.
The math behind frontier models is akin to this kind of discovery. Data science, linear algebra, information theory, physics…these are among the elements that feed into model development. And they’re all available to anyone (i.e., academic journals, public code repositories, etc.) even if not everyone would be able to understand them. We’d be, if even unintentionally, attempting to regulate “science” in a way that (1) we’ve never done before, for good reasons, and (2) could foreclose some of the most exciting and beneficial discoveries in human history.
Claim #4
AI models are general-purpose tools that only cause harm through a user’s malicious application of them, and existing law already governs that nexus of bad actor and bad act.
Technology is a tool–it is neither good nor bad. Uses are what we should be focused on. ChatGPT, Claude, et al. are general purpose in nature. They don’t intrinsically present a harm - new in type or otherwise. Like an internal combustion engine, which can be used to inflict horrific harm on others by a driver when assembled together with other parts into a car, the model itself is dependent on a dangerous user and harmful application to inflict harm.4
There is risk of misuse, and that concern is legitimate and deserves to be taken seriously. But that is a different consideration than attempting to reduce harms via development regulation. Even outputs that might appear suspect on the surface can be used by well-meaning users to prevent harm, by better understanding for instance cybersecurity vulnerabilities. It’s at the nexus of a malicious user and execution of a discrete act where laws are best understood to be appropriate.
Sub-Claim
Existing laws do more of the necessary work to deter harmful acts that involve AI than you think. The vectors of harm are not fundamentally changed by the existence of AI. Employing AI to commit fraud, theft, assault, or any other form of illegally harming another person are and should be equally prosecutable under the law as though they were committed by other means. There may be some tweaking of current law (consumer protection and so on) to ensure these acts are not excluded for technical legal reasons, but this is categorically different from adoption of a whole new legal regime allegedly addressing novel protections.
Claim #5
A fragmented, fast-growing patchwork of state AI laws is a more immediate threat to innovation than any AI harm itself, and only federal preemption can resolve the underlying collective action problem.
The most immediate and salient policy risk we should be concerned about is unrestricted proliferation of new state laws. Beyond the reasons stated above, even the best intentioned and knowledgeable legislatures leads to an unworkable morass of laws in the near- to medium-term. No innovators will be able to ensure compliance with all of them–already there are some 1,200+ bills pending in state capitals.5
The interest from state legislators is understandable, but there’s a collective action problem. Even broadly aligned policy goals between states will be undermined by the lack of incentive to reconcile granular details which may appear minor, but will result in significant regulatory costs when applied broadly. This is not even to speak of whether the laws themselves are well-designed.
The nature of software, the internet, and modern communication modes means there’s no practical way to withdraw from particular state jurisdictions where the legal structure is unworkable. An ill-informed law in one state becomes a problem for everyone.
We are at the takeoff phase of this incredible technology, but we risk throwing debris onto the runway if we do not address this problem. Which is why a federal law is necessary to solve the commons problem by preempting state AI law. It’s the only way to rationalize policy at a critical moment for this emergent technology.
I welcome thoughts on how I did.
I’ve omitted some other recurring arguments in the above direction, mostly for space and because I find them the least compelling. For example:
Definitions and other technical elements of these laws, like the CA one, even for their own intent don’t capture the relevant aspects they want to. They are not informed by the most sophisticated practioners and end up using irrelevant, naive metrics. Further they are inflexible to a dynamic frontier.6
Regulation, especially targeted at the model development layer, will only serve to support large incumbent entities, at the expense of newer, smaller, arguably more innovative startups. Regulatory capture takes hold, and the incidence of fixed regulatory costs disproportionately hits smaller innovative companies.7
I coudn’t bring myself to attempt an earnest defense of these because anything beyond a sentence or two become so obviously obtuse it felt like an unfair handicap to this exercise.8
ACTUAL PRIORS:
My view has for some time been the following:
The long-term effect of AI advancement and proliferation will be on net positive.
This will include benefits (widely defined) outweighing costs.
Some of those costs are/will be nonetheless addressable.
Thus we should consider some regulatory/governance structure via law that places some binding constraints on frontier model labs and downstream uses.
Therefore, I strongly oppose a 10-year moratorium on new AI laws.
Broadly: Like other technological advances, the long-term equilibrium will be neither catastrophic nor utopian (though non-zero possibility of either).
Cf. Marc Andreessen, “Why AI Will Save the World,” a16z, June 6, 2023: “Historically, every new technology that matters… has sparked a moral panic.” Andreessen argues such panics have consistently proven overblown relative to the long-run gains delivered by the technologies in question. https://a16z.com/ai-will-save-the-world/
Cf. Adam Thierer, Permissionless Innovation: The Continuing Case for Comprehensive Technological Freedom (Mercatus Center, 2014/2016), framing the choice as “permissionless innovation” versus the “precautionary principle.” On the recombinant DNA/Asilomar sequence specifically: the 1975 Asilomar Conference produced voluntary containment guidelines that NIH later formalized and progressively relaxed as research demonstrated lower-than-feared risk; see National Research Council, Biotechnology Research in an Age of Terrorism (2004), for a retrospective on the NIH Guidelines’ evolution. Recombinant human insulin (Genentech/Eli Lilly’s Humulin, approved 1982) was the first recombinant-DNA drug to reach the market.
Cf. Matt Perault, “Regulate AI Use, Not AI Development,” 2025, and Marc Andreessen, “Why AI Will Save the World” (2023), both of which frame frontier model training as continuous with, or substantively akin to, basic scientific research that has historically gone unregulated.
Cf. Andreessen, “Why AI Will Save the World” (2023): “Technology is a tool. Tools, starting with fire and rocks, can be used to do good things…and bad things.” https://pmarca.substack.com/p/why-ai-will-save-the-world
Cf. Kevin Frazier & Adam Thierer, “1,000 AI Bills: Time for Congress to Get Serious About Preemption,” Lawfare, May 9, 2025, arguing the state-by-state proliferation of bills (a milestone reached within four months of 2025) risks an incoherent patchwork undermining national AI competitiveness. https://www.lawfaremedia.org/article/1-000-ai-bills--time-for-congress-to-get-serious-about-preemption
https://www.cognitiverevolution.ai/a16z-on-protecting-little-tech-the-techno-optimist-ai-policy-agenda-with-matt-perault/
Ibid. (But also everywhere).
Related, but not relevant, many of these talking points when heard directly from the sources, can’t help but betray how newly discovered they are — reeking of the high school student who just read Catcher in the Rye then makes it his whole personality.