Streetwise Professor

AI hysteria has predictably led to calls to regulate! regulate! regulate!, seemingly from all corners except for the Trump administration and NVIDIA’s Jensen Huang. Those baying for “regulation” include frontier model leaders Anthropic and OpenAI.

I put “regulation” in quotes because as is often the case, those calling for regulation are imprecise on just what regulation should entail, and have evidently given little serious consideration to how regulation works in practice. The word is usually more of a magic incantation than a description of something real that has any chance of realizing the goals of those using it.

So what should AI regulation look like? To answer that question, it’s important to start at the basics.

Decades ago Steve Shavell wrote two important articles that speak to the subject: “Liability for Harm Versus Regulation of Safety” (JLE, 1984) and “The Optimal Structure of Law Enforcement” (JLS, 1993). Shavell laid out several criteria for determining the best way to regulate.

Broadly speaking, there are two alternative methods of regulation. Harm-based (“ex post“) sanctions that attempt to deter harmful conduct, and preventative (“ex ante“) regulations. The former includes the common law tort system, and civil and criminal laws. The latter includes things like government approval for the sale of a product, with FDA approval of drugs being an example. Many of the calls to regulate AI fall in the latter category, and some members of the Trump administration basically advocate the former.

So which is best here?

Shavell identifies several factors that determine the optimal legal structure. On balance, these favor ex post/harm-based tort-like approaches.

The first criterion is “is the possibility of a difference in knowledge about risky activities as between private parties and a regulatory authority.” That is, information asymmetry between the regulator and the regulated.

This is quite intuitive. If an actor (an AI firm, in this instance) has better information about the likelihood its actions will cause harm, and the amount of harm it will cause than does a regulator, if it knows it will pay damages commensurate with the harm it causes, it can take the efficient precautions. It will appropriately trade off the cost of mitigating harms (including the cost associated with reduced performance) with the cost of the harms because the prospect of paying damages based on costs and probabilities of harms will induce it to internalize the costs of its actions.

In contrast, an ignorant regulator knows less about the probabilities and magnitudes, and is therefore likely to implement measures that are either overinclusive or underinclusive. That is, regulators are far more prone to Type I and Type II errors.

It is clearly the case with AI that the developers-the labs/frontier firms in particular-will be far better informed about their creations and the risks that they pose than any government regulator would be.

Shavell’s second criterion is “that private parties might be incapable of paying for the full magnitude of harm done.” That is, in the present context, an AI firm might be “judgment proof.” In the event, the actor will not internalize the costs it imposes on third parties, and will tend to take excessive risks and impose excessive harms. Here, external interventions that constrain actions in order to reduce the magnitude and frequency of harms can be superior to deterrence.

The inability to impose monetary damages commensurate with harms is one justification for criminal sanctions, including imprisonment or even execution. These are non-monetary penalties that enhance deterrence.

This is one factor that might-might-favor ex ante regulation. Anthropic or OpenAI (or any other firm) could not pay damages commensurate with harm if that harm is the end of humanity!

But this brings us back to the first issue. In deciding whether judgment proofness is indeed a problem, we need to have some assessment of the distribution of harms, and here the AI firms have a decisive information advantage. And as I discuss below, we absolutely cannot take at face value the pronouncements of Amodei, Altman, Musk, et al, on the potential harms posed by AI, or their likelihood.

And more, even though judgment proofness may preclude AI firms from fully internalizing costs of their products (especially the harms they may create), bureaucrats operating subject to low powered incentives clearly don’t internalize them. Nor do they internalize the benefits of more capable AI. Therefore, although judgment proofness may be a necessary condition for ex ante regulation, it is not a sufficient one.

Shavell’s third factor is “the chance that parties would not face threat of suit for the harms done.”

In the US legal system? As if! Indeed, the likely bigger problem with relying on the tort system to deter would lead to suits for harms not done, i.e., frivolous or baseless litigation.

Shavell’s last factor is the relative administrative costs of the two systems, i.e., the cost of a regulatory bureaucracy vs. the costs of litigation.

Both litigation and bureaucracy are costly. The advantage of ex post, harm-based regulation is that the litigation costs are only incurred when a harm incurs: no harm, no foul, no legal expense (with the caveat related to frivolous litigation mentioned above). In contrast, the bureaucracy must be fed continuously, even in the absence of harms.

And note that regulation would necessarily have to be international in scope. Look at the post-Great Financial Crisis history of banking and financial regulation for a demonstration of the monumental costs and challenges of international regulation. AI regulation would put even that in the shade, especially since it adds China and strategic competition into the mix.

To which, I would add another factor: influence costs, including regulatory capture. It is well known that the regulated exploit their information advantages and financial resources to induce regulators to take actions that benefit some regulated entities, at the expense of other entities and the public at large. This is especially likely to work when the regulated have a big information advantage (check) and a lot of money (check).

Overall, to me the scorecard tilts heavily in favor of ex post, harm-based, liability mechanisms as opposed to ex ante regulation. This is especially true given the radical uncertainty and information asymmetries inherent in such a dramatically new technology. We are learning the harms as we go, and the developers learn more and faster than any bureaucrat can. Indeed, being liable for harms gives the developers a strong incentive to identify those potential harms-their magnitudes and likelihoods. A far stronger incentive than bureaucrats have.

And that provides a segue into a discussion of one of the most fascinating aspects of the ongoing debate/hysteria: the fact that the frontier labs are the loudest and most strident voices calling for regulation. In my opinion, this is overdetermined.

A major reason is that Anthropic et al are deathly afraid of being liable for the harms their products cause. Palantir’s Alex Karp perhaps overstates the case when he says they are effectively calling for nationalization in order to avoid liability, but I think it’s pretty clear that avoiding paying for the harms their products cause is a major driver behind the calls–pleas, really–for regulation.

Which, to be honest, another reason to favor a liability-based system. It suggests they understand that even precluding the end of the world or singularity scenarios, they are aware that their products can do great damage. The only way to get them to act on this knowledge is to force them to pay the costs, and not be able to say “don’t look at me! The regulators approved what I did!”

Another reason-the one that has received most commentary-is that the frontier firms know that they would be able to capture their regulator(s) by exploiting their information advantage and vast wealth. And as George Stigler pointed out long ago, the biggest benefit a regulator can bestow on its industry masters is a reduction in competition, including by the erection of entry barriers. As I’ve written with respect to financial regulation, regulatory costs are largely fixed and therefore benefit large incumbents at the expense of new, smaller, disruptor entrants.

There is another reason that is laid bare by some of the specific rhetoric and actions of Amodei, Altman, et al: a belief that intense competition between the frontier firms, let alone between them and open weight models, small models, etc., will be ruinous to them.

The cabal has been negotiating in public for “pacing,” i.e., a coordinated slowdown of model development. Competition on the capability dimension is hugely expensive, especially with respect to training. I view the various public statements-including the apocalyptic rhetoric justifying a slowdown-to be “cheap talk” intended to coordinate reduced investment. A way of cartelizing the industry.

Which leads to another motive for favoring regulation–protection from antitrust laws.

A coordinated slowdown would also address something else that I think looms over them. The whole AI sector is a bet on future demand, future prices (which depend on demand and competition from non-frontier implementations), the costs of production (e.g., data center investment and operational costs, including notably electricity), and the execution timeline (e.g., will bringing data centers online be delayed by supply chain bottlenecks or local political opposition?) Doubts have been increasing on all these dimensions, and I strongly suspect that the industry insiders fear that their massive investments (at all levels of the value chain) will not pay off if the current pace and growth are maintained. That is, the “pause” is not driven by safety fears, but by fears of a financial bloodbath.

It reminds me of Stalin’s pausing collectivization because he was “dizzy with success.” AI has been so successful that it has led to dizzying excesses that need to be reined in.

So there are many reasons for AI executives to want a pause. I think all are operative, and all push in the same direction–regulation by means other than liability.

But the nature of the industry and the product and the economics of harm reduction mean that a liability mechanism is the best way to address the safety concerns surrounding AI. That mechanism would give AI developers the strongest incentive to control the risks that they lecture us about daily. Anthropic doesn’t need OpenAI to mitigate the risks of its models, and vice versa. Internalize costs via a liability system and they will unilaterally have the incentive to do so.