What Overruling Chevron Means for AI

Loper Bright v. Raimondo and the future of AI policy

I said I wouldn’t write this week and look at me here, writing. Go figure! Happy July 4^th, everyone.

In the aftermath of the Civil War, New England had a curious problem: fish populations along the coast were dwindling. As a result, Congress created the United States Office of Commissioner of Fish and Fisheries, the oldest environmental conservation and regulatory agency in America. Its job was to study this and other environmental problems related to fish populations in our territorial waters. It had one employee, and a budget of $5,000.

Today it has 4,200 employees and a budget of a little more than $1 billion. It has since been renamed the National Marine Fisheries Service. Among other things, it is responsible for implementation of something called the Magnus-Stevens Fishery Conservation and Management Act, a law passed in 1976. The purpose of that law is to regulate overfishing, primarily by foreign ships operating just outside America’s territorial waters. It did so by extending America’s maritime borders from 12 miles off the coast to 200.

The NMFS was put in charge of administering this vast new territory—at least as it pertained to the fish population that lived within it. One of the ways it accomplished, and accomplishes, this goal is to place observers on ships to ensure that they are obeying the law. Sometimes, for example if your ship is not American or if you are fishing off the coast of Alaska, the law requires fishermen to pay for these observers while they are aboard (their time, food, etc.). But most of the time, the NMFS pays for the observers and you, as the fisherman, simply have to let them board your vessel.

That is, until 2013. Facing budgetary constraints, the NMFS decided that it would start making American fishermen in places other than Alaska start paying for the fish police to watch them fish. Nothing about the law had changed, so on paper, the agency had no legal basis for this change. The agency justified its decision under a judicial doctrine called Chevron deference, which came out of a 1984 case in which the US Supreme Court decided that if an agency’s interpretation of existing law was “reasonable,” courts could not dispute that interpretation—even if the judges’ own interpretation meaningfully differed from the agency’s.

It turns out (as we will re-discover if California’s proposed AI regulation, SB 1047, passes) that the word “reasonable” is bigger than it looks. Chevron deference has been used to radically expand NIMBY mainstay laws such as the Endangered Species Act and the National Environmental Policy Act. It has helped to expand the power of essentially all federal agencies. It’s been used extensively in the regulation of technology, too: the Federal Aviation Administration’s regulation of drones, the Federal Communication Commission’s Net Neutrality regulations (and the FCC’s ability to undo those regulations under the Trump administration), and the Department of Transportation’s regulation of autonomous vehicles all rely, either implicitly or explicitly, on Chevron deference.

A lot of people think this is unfair, at least as it applies to their particular niche. One of them is a New Jersey-based fisherman named Bill Bright, the owner of Loper Bright Enterprises, who objected to having to pay for fishing observers on his boat.

In Loper Bright Enterprises, Inc. v. Raimondo (and yes, that’s Gina Raimondo, the Secretary of Commerce; the NMFS is housed within the Commerce Department), decided last Friday, the Supreme Court put a permanent end to any future use of Chevron deference. Cases decided in the past on the basis of Chevron deference are still in effect—but they too may come under new scrutiny. The Supreme Court has reclaimed what it views as a fundamental judicial responsibility: serving as the final arbiter for what constitutes a “reasonable” interpretation of existing law. As Chief Justice John Roberts, author of the majority opinion, put it (no emphasis added):

… agencies have no special competence in resolving statutory ambiguities. Courts do. The Framers anticipated that courts would often confront statutory ambiguities and expected that courts would resolve them by exercising independent legal judgment. Chevron gravely erred in concluding that the inquiry is fundamentally different just because an administrative interpretation is in play. The very point of the traditional tools of statutory construction is to resolve statutory ambiguities. That is no less true when the ambiguity is about the scope of an agency’s own power—perhaps the occasion on which abdication in favor of the agency is least appropriate.

On its merits, it seems to me that the Court’s decision in Loper Bright is correct. Nonetheless, I think the effect of this decision on the regulation of AI and other emerging technologies will be more complicated.  

First, a macro point, not specific to technology policy: Chevron deference has been diminishing in importance for the past decade. The Supreme Court has not used it to decide a case since 2016. And as mentioned above, it only applies to decisions going forward—no prior judicial decisions are reversed because of the ruling in Loper Bright. So in general, the effect of the ruling will take years to become apparent and may be smaller than some are forecasting.

Second, a micro point, for those of you hopeful (or worried) that this decision may affect SB 1047 and its proposed regulator, the Frontier Model Division, I have bad (or good) news: Chevron deference has little to no bearing here, because it pertains to federal agency interpretations of federal laws, and SB 1047 is a state law.

That’s not to say there is no effect on AI policy more broadly, though. If you’ve been reading my work for a while, you know that I favor a conduct-based approach to AI regulation: our existing laws, I believe, codify the standards for conduct “we” wish to see in the world. Thus, those laws are a reasonable basis on which to build “AI policy.” This is basically what happened with some of the examples I mentioned above: I think it is reasonable for the DOT to regulate self-driving cars, the FAA to regulate drones (and related software to enable drone autonomy), and the FDA to regulate AI models in medical devices.

Loper Bright does not exclude the possibility of similar outcomes in the future. But it may make them more difficult to achieve. Agencies will still be able to issue their own interpretations of existing law, it’s just that now, a judge will be able to exercise more discretion over whether that interpretation is correct. Even in an ideal judicial climate, this would likely mean that laws will be interpreted unevenly and inconsistently. Chief Justice Roberts admits as much in the majority opinion: “We see no reason to presume that Congress prefers uniformity for uniformity’s sake over the correct interpretation of the laws it enacts.”

And with technology policy, we live far from an ideal judicial climate. Indeed, just a few days before issuing the decision in Loper Bright, the Court punted yet again on a major internet case—this one focusing on the government’s efforts to censor speech on the internet (in this case, the Court punted by claiming that the party did not have the right—standing—to sue). Last term, the Court similarly punted on two major social media cases. Another two cases, pertaining to social media regulations issued by Texas and Florida, were decided today, and the Court seems to have punted once more (Update: maybe not. I’ll cover the NetChoice decisions in detail in a follow up post).

It would be one thing if courts demonstrated a keen ability to issue wise judgments regarding internet policy, but again and again, we have seen them return null on dire issues—like how government regulation of social media interacts with the First Amendment. Instead, they find reasons not to issue clear decisions: Supreme Court Justices “are not the 9 greatest experts on the internet,” Justice Elena Kagan said during oral argument for one of the punted social media cases last term.

It strikes me as unreasonable for Chief Justice Roberts to talk a big game about the judiciary’s boundless capacity for resolving legal ambiguities, even on highly technical matters, while his court is consistently unwilling to… resolve legal ambiguities on highly technical matters pertaining to the internet. This does not bode well for the regulation of AI, which could become far thornier than the regulation of social media platforms.

Indeed, Justice Kagan, who dissented from the majority opinion in Loper Bright, feels similarly (emphasis added):

[The majority decision] gives courts the power to make all manner of scientific and technical judgments. It gives courts the power to make all manner of policy calls, including about how to weigh competing goods and values. It puts courts at the apex of the administrative process as to every conceivable subject—because there are always gaps and ambiguities in regulatory statutes, and often of great import. What actions can be taken to address climate change or other environmental challenges? What will the Nation’s health-care system look like in the coming decades? Or the financial or transportation systems? What rules are going to constrain the development of A.I.? In every sphere of current or future federal regulation, expect courts from now on to play a commanding role… It is a role this Court has now claimed for itself, as well as for other judges.

It is unfortunate that Justice Kagan chose the word “constrain” here, because there need not be anything constraining about the application of existing laws to AI. Indeed, what is far more likely to constrain AI and other emerging technologies are sweeping new laws written with those technologies in mind.

Things could play out differently. Adam Thierer, for example, intelligently observes that much of present-day “AI policy” (for instance, the kind provoked by President Biden’s Executive Order) is promulgated to avoid Chevron issues altogether. Chevron deference applies to agency rulemaking under a law called the Administrative Procedure Act, but agencies can also issue so-called “soft law”—"guidance,” “advisory memos,” “recommendations,” and other mechanisms that avoid the specific agency rulemaking process over which the Loper Bright decision obtains. Instead of handing power to the courts, then, agencies could simply opt to use these even more esoteric and nebulous regulatory tools.

To wit: the Biden Executive Order on AI uses the Defense Production Act to require AI companies training models above 10^26 flops to report basic things about their training run and safety practices to the Department of Commerce. The DPA unambiguously allows this per se, and as a general matter, I support basic reporting at the frontier of AI. But accomplishing that policy goal as President Biden did means, fundamentally, that regulation of frontier AI models is taking place via emails between AI companies and Department of Commerce employees, and not as part of any formal legal process.

What if the Department of Commerce says, “we don’t like this safety practice, you should do something else?” Or what if they say, “we think you should wait to release this model until after the election”? No one, other than the AI company in question, would have standing to sue, and it’s not entirely clear (to me, at least) on what grounds they could sue. At the very least, such “soft law” is far from the reach of Chevron deference or the Court’s decision to overturn it.

And governance of this kind—if it can even be called governance—may be outside the bounds of the law entirely. One of the social media cases the court punted this term focuses on government employees informally asking social media companies to censor certain forms of content. The Court’s head-scratching decision that no one involved had legal standing suggests that “governance by email” may be applied more aggressively in the future. This is undoubtedly worse than Chevron. Rather than driving policy back into the halls of Congress, decisions like Loper Bright could have the unintended consequence of driving bureaucracies further into the mountains in their quest to escape the bounds of the rule of law.  

Another possibility is that Congress begins writing laws to explicitly grant agencies more leeway in their interpretations. Importantly, the majority opinion in Loper Bright did not overrule Chevron on Constitutional grounds: it merely said that Chevron is inconsistent with the Administrative Procedure Act. Though Justices Gorsuch and Thomas did argue that Chevron is unconstitutional, the majority opinion did not. For the time being, then, there is nothing that would stop Congress from simply writing Chevron deference into current or future law.

I would be hesitant to support such a move by Congress, because in the grand scheme of things, I believe Loper Bright is the correct decision: courts really should resolve statutory ambiguities, even if they are highly technical. Anything else is a formula for unchecked growth in unaccountable administrative agencies—as the four decades since Chevron have made clear.

More than anything, then, the Loper Bright decision makes me hope that the Supreme Court will start issuing clearer judgments about internet-related law. I’m not holding my breath, though, and that’s a shame. More uncertainty is the last thing we need. It may well be what we get.