Who Should Regulate What?

I wanted to remark on this Washington Post editorial.

On Thursday, President Donald Trump abruptly pulled the plug on the signing of an executive order that would have given federal agencies an early look at the nation’s most powerful artificial intelligence models before their public release. That’s good news: Even the supposedly voluntary review system under consideration could have hardened into a government chokepoint on U.S. AI development.

Here are some additional significant snippets:

The order would reportedly have created a framework for leading AI labs to share their most advanced models with a coalition of national security agencies before public release. It would also have established a joint cybersecurity clearinghouse, run by both government agencies and private-sector representatives, tasked with finding and patching vulnerabilities in the nation’s digital infrastructure before adversaries could exploit them.

and

A coalition of more than 60 Trump allies, organized under the banner of Humans First and backed by Steve Bannon, wrote the president urging mandatory government approval of frontier AI systems before release. Kevin Hassett, the director of the National Economic Council, publicly floated an FDA-style regime earlier this month in which AI models would have to be “proven safe” before they could reach the public.

and

AI poses real cybersecurity risks, but they need to be addressed with care and precision. The president is right to be wary of creating a bureaucratic review regime that erodes America’s edge.

I’m going to divide my remarks into four sections:

• The problem
• The urgency
• What the creators don’t understand
• The solution

The problem

The snippets above call out the national security implications and the “safety” implications. In addition to those there are infrastructure issues and liability issues. The infrastructure issues are highlighted by this graph:

U. S. consumption increased rapidly from 1970 to about 2000, plateaued due to various efficiencies realized, and, driven largely by generative AI and electric vehicles, began to increase around 2020. The rapid build-out of data centers and increased adoption of EVs will ensure that growth continues for the foreseeable future.

At some point utilization will exceed the grid’s capacity. That is likely to produce a series of effects. First, producers may implement controlled voltage reductions of 5–8% to prevent cascading failures. Historically, that has frequently been followed by “rolling blackouts”. Together those will produce wear-and-tear on devices containing electric motors, such as refrigerators, air conditioners, and well pumps. The costs of these effects will not be borne by commercial consumers whose contracts ensure continuous use. They will be borne by household consumers.

Neither executive agencies nor Congress possess the technical knowledge or adaptive speed necessary for effective ex ante control of frontier AI systems. We are in terra incognita. The machine learning researchers who build these systems understand the architecture and the training process. They do not understand and have said publicly they do not understand what the systems have learned, why they produce the outputs they produce, or what they will do in novel situations. That is an accurate description of the current state of knowledge.

The urgency

Basically, we can’t wait for studies or well-formed regulations. The first lawsuits against gAI companies are already being filed. Regional power constraints associated with accelerated data-center growth are already emerging in several U.S. markets, particularly in Texas where data centers are being built. The pace at which data centers are being built is considerable.

What the creators don’t understand

The enormous irony in this is that not only do elected officials or any foreseeable regulatory agency lack the knowledge and understanding, the creators of gAI don’t have the understanding, either. I’ll flesh that out in a later post.

The solution

Fortunately, there is a prospective solution but it’s one I believe with be very controversial and result in considerable backlash: strict liability. Under a regime of strict liability those harmed by gAI would not need to show intent or negligence on the part of the AI companies—all they would need to show are involvement of the companies and harm.

The response of the companies and, presumably, some economists would in all likelihood be that would have a chilling effect. That is an alarmist claim. Strict liability would force the industry to price risk realistically rather than externalizing it. Strict liability would alter incentives but that is precisely its purpose. The companies could continue to proceed at their own pace. It would only ensure that the risks of their actions would be borne by those who benefit most—the companies rather than those who benefit least—ordinary household users and consumers of electricity.