For the better part of the year, artificial intelligence (AI) policy has, reasonably, focused on establishing guardrails to mitigate some of the biggest risks in developing and deploying frontier AI models. This year has seen a rapid expansion in model capabilities, with the release of the first models capable of conducting advanced cyberattacks. We have also witnessed the first instances of models becoming advanced enough to outsmart the humans stress-testing them, raising some concerns about potential loss of control over frontier AI. In response, policymakers have enacted measures to mitigate some of the risks associated with the development and use of these advanced models. From the White House evaluation framework to the Joint Commitment on Frontier Responsibilities, rules setting up the AI industry’s duties and responsibilities have progressively increased.
However, while the industry faces increasingly intense scrutiny, there is worrisome vagueness about the government’s role and authority as a regulator.
Earlier this year, the administration enacted export controls on one of the leading frontier AI companies with little to no justification. Emil Michael, the vice chair of the White House’s “Super Intelligence Task Force,” has said it will likely levy them again against a company that does not “cooperate” with the government. In fact, he claims the executive has a wide range of tools at its disposal.
At the same time, reports indicate that the Federal Trade Commission is already investigating OpenAI over the loss-of-control incident that led to the Hugging Face hack, an investigation that is running parallel to its intent to pursue action against AI companies if it considers any ideological bias to be a deceptive practice under Section 5 of the FTC Act.
In summary, while the industry and the public are seeing an influx of rules and agreements charting a path forward for AI companies, the government has not seen the same. There is no clarity regarding which government agency or entity is responsible for evaluating the industry’s adherence to these commitments. There are no thresholds that clearly delineate when a company has (or has not) failed to live up to the bargain. In fact, even though a lot of these agreements are a byproduct of a nominally voluntary negotiation, the statement by Michael could be interpreted as another instance of jawboning, where companies are forced to sit at the table and reach an agreement, or they would be subject to a nondisclosed and potentially ever-expanding set of sanctions by the executive branch. To make it worse, in the instances in which the government claims to limit its power, such as by avoiding a pre-release vetting regime, it is creating de facto avenues to curtail such commitment to retain an ability to prevent AI models from being released or to pull them out of the market when it wants to.
The risks of granting the government wide and largely unchecked latitude over the AI industry should not be brushed off lightly. As my colleague Jennifer Huddleston and I have argued, the development, deployment, and use of AI are embedded in a vast array of expressive behaviors that are protected by the First Amendment. While the riskiest, most militaristic uses of AI have been top of mind during these discussions—a reasonable response, given the immediate risks—the expressive behaviors commonly associated with civilian use of AI must not be disregarded. As there are no current guidelines or statutes that specifically limit government action to narrowly target riskier uses, this and any future administration could easily invoke these regulatory powers to sanction an AI company that it considers “too woke,” “biased,” or spreading misinformation or disinformation.
As of right now, the executive can levy significant sanctions on an AI company without much justification, as it is not bound by a specific statute or a set of predetermined rules. It does not need to publish its reasoning to the public; there is no recourse for appeal, as it is not required to be tied to a specific set of criteria, and it has no threshold it must clear before it can exercise its power. It can unilaterally deem a model dangerous or unfit for release, and the citizenry must take its word for it. It is easy to see how this power could be abused by a president who merely wants to antagonize a company by claiming it is not sufficiently cooperating with the government or that its product is risky in order to punish it for constitutionally protected expressive behavior. Because the executive does not need to publish its reasoning, the company has no way to turn to the courts to scrutinize the government’s action.
Michael’s words should serve as an alarm for both industry and policymakers. While setting guardrails to mitigate the risks associated with frontier AI systems is important, defining the government’s role and power needs to be addressed with similar urgency. The current regulatory approach to AI gives the government wide latitude to intervene in and regulate the AI industry, with few checks on its power to protect citizens from the weaponization of that power for censorious ends.
Any sitting president could take a model out of the market or prevent it from release on a whim, and they do not need much justification to do so. This opens the door for cronyism, government abuse, and an erosion of the rule of law. Setting guardrails against government abuse must be as important as the guardrails against AI-related risks.
