After the recent midterm elections, it seems that the incoming Congress will come with a revitalized, ambitious appetite to pass legislation on artificial intelligence (AI). As my colleagues have highlighted, some in Congress have called for aggressive regulation of the industry, even proposing pauses on development. But the political reality is that any proposal from the new Congress will need to survive a presidential veto from an administration that has expressed a desire to maintain a light-touch approach to AI regulation. This equilibrium will, fortunately, fend off proposals for heavy-handed regulation—opening up a unique opportunity to address some of the concerns surrounding the AI industry without hampering the development of one of the most promising technologies in decades.
Focusing on narrowly targeted, restrained legislation is Congress’s best option to address some voters’ concerns about AI. But Congress’s attention should look beyond regulating the AI industry towards curbing executive power. The executive’s notorious overreach highlights the need to create a more orderly, rules-based system for AI governance. Policymakers should be wary of passing legislation that could grant the executive more power over AI.
On Frontier AI Safety
Take, for example, a potential frontier AI safety bill. Any attempt at enacting an AI safety bill faces a significant, almost paradoxical, challenge: It needs to appropriately tackle any risks associated with the development and deployment of frontier AI models without giving the government the ability to arbitrarily wield power to shut down a model owned or developed by a company it does not like.
As we have learned from the White House’s application of its evaluation framework, granting excessive discretion to the executive will likely result in that power being used arbitrarily and unfairly. Any proposal needs to balance immediate safety risks with the need to embed robust protections against unchecked executive power. These checks should not be taken lightly and must be enacted to prevent the potential weaponization of executive power by whoever is sitting in the Oval Office.
Giving the government wide latitude to shut down a model might be an efficient way to thwart a safety incident. Still, it does so at the cost of creating avenues for the government to censor AI companies whose models produce content it dislikes. On the flip side, strict safeguarding of free speech rights for AI companies and their users will likely introduce friction, slowing the government’s ability to respond to a safety incident.
This tradeoff is a byproduct of AI’s nature as a dual-use technology. In the same way it can be used in riskier, militaristic applications such as cyberattacks or bioweapons development, civilians can use it to conduct research or run a political campaign. Unless regulation can draw a clear line and narrowly target its riskiest applications, any AI regulation will be forced to strike a balance between speech and safety.
One way to narrow the scope of such a bill is to establish a “catastrophic risk” threshold for executive action that is readable, largely objective, and clear-cut. This was the approach attempted in Anthropic’s recent discussion draft that it pitched to lawmakers. It defined catastrophic risk as “materially contribut[ing] to the death of, or serious injury to, more than 50 people, or to more than $1,000,000,000 in damage to or loss of property.” Exactly where to draw the line on the threshold might warrant further debate, but the importance of the clause is that the threshold that triggers government action should be as clear-cut and black-and-white as possible. Gray areas leave room for the government to wield this power arbitrarily and unfairly.
Similarly, attempts to shut down a model must be argued and defended before a judge. When the government burdens constitutionally protected rights, such as through searches and seizures or restraints on speech, it usually requires judicial review and authorization before proceeding, even when national security is at stake. For scenarios of imminent risk where a timely response is a priority, we usually create avenues for expedited review, but we do not—and should not—remove the requirement of judicial review. Judges are best equipped to balance safety against speech rights. The executive, by contrast, has little incentive to do so, as its recent treatment of Anthropic shows.
No member of the executive branch should be able to make the unilateral decision to shut down a model without having to plead the case to a judge who can properly evaluate the merits of that decision. The agency responsible for making that decision should be explicitly designated to avoid confusion or interagency conflicts. The criteria the government needs to meet to prove its case should be explicit and clear-cut to prevent ambiguity that could be exploited to expand its power beyond its original intent. And the onus should be on the government to prove its case, not on a developer to prove a negative.
Another policy alternative to narrowly target harmful behavior could be to clear the uncertainty regarding the application of tort law to AI agents, especially while they are in development. As some have pointed out, the current tort law system is likely incapable of addressing harms that could stem from rogue agents not commercialized to the public, such as the internal agents involved in the OpenAI-Hugging Face incident. Under current tort law, the vicarious liability that companies face when human employees commit harms while pursuing a business interest does not apply when an agent commits those same harms. Ultimately, the goal should be for companies to internalize the costs of agent misbehavior that are currently unpunished under our legal system, thereby incentivizing them to ramp up their safety and containment procedures. Others have suggested introducing a duty-of-care standard for AI development, which could be another way to internalize these costs.
There Is No Need for a Data Center Moratorium, But There Should Be Better Information
A major point of contention during this election cycle was the rising negative sentiment against data centers. As my colleagues have pointed out, this animosity toward data centers has also led to calls for national moratoria on new data center construction. As they have also noted, the negative sentiment has been largely driven by narratives about excessive power or water consumption that do not align with reality.
Rather than imposing moratoria and overriding decisions that localities themselves should make, Congress can ensure greater transparency in these deliberative processes and, more importantly, enable honest, informed discussions that allow citizens and policymakers to determine whether constructing a data center is a good deal for their communities.
One way to do that could be through a policy that tasks the government with building a national database and establishing a unified national reporting standard for energy and water consumption, noise, and pollution statistics. This would give policymakers and advocates a resource to appropriately assess the local impacts of data center construction and properly compare them with other industrial uses of land and resources. Some labs, like OpenAI, have already suggested a similar measure at the state level, but having a federal reporting standard might be the most effective way to streamline this conversation.
Federal Preemption Remains Urgent
As I have argued before, addressing the patchwork of AI governance laws should be an urgent matter. Digital technologies have shown that local regulations often have costly, confusing spillover effects on businesses and consumers alike and are self-defeating. Some proposals, such as Anthropic’s draft, have introduced a narrower form of preemption that preempts only issues explicitly regulated by a bill. This means that states would retain the authority to regulate unaddressed issues as a compromise to address opponents’ qualms about federal preemption. Broad preemption remains the most effective way to ameliorate the confusion caused by conflicting regulatory standards, but a narrow preemption that targets the most obviously interstate-commerce-level of the technological stack, such as AI development, is preferable to the current status quo of letting state regulation run amok.
Conclusion
The resounding Democratic win in the midterm elections seems to have emboldened some of the groups pursuing aggressive regulation of the AI industry. However, the resulting balance of power between Congress and the executive branch dictates that a lighter-touch, narrowly targeted approach is the most likely to become law. Conveniently, this approach is most likely to quell some of the current concerns surrounding the technology without unduly hampering its development.
