Everyone in Silicon Valley is talking about antitrust law. The main question being asked is this: will American antitrust laws stop frontier AI labs from cooperating with each other on safety, risking catastrophe? The potential culprit, the argument goes, is the Sherman Act, an 1890 law that prohibits monopolies and anti-competitive business practices. What this conversation misses is that the law already allows for an exception that would be applicable here. It’s called Parker immunity, named after the related Supreme Court case.
But, before we get to the case law, we need to understand what the fuss is all about. For the last several years, the leading American AI companies have been in a fierce competition against one another. They each want to produce the most capable AI model in the world. Now, a string of incidents in which AI models broke into systems and government websites they were never meant to touch has shown just how capable these models have become. As a result, some AI companies are singing a different tune.
Dario Amodei, the CEO of Anthropic, wants to cooperate with others to slow down—or in his words, “pace”—AI development. He argues that the technology has improved too quickly to be able to meaningfully test it and ensure its safety. But Amodei believes he cannot stop developing the technology unilaterally (lest OpenAI beats Anthropic and builds the most powerful AI). Instead, he wants to collaborate with his competitors. However, when competitors agree to ease off the competitive pressure, their cooperation starts to look like collusion, which the Sherman Act prohibits. So, Amodei, and many others in Silicon Valley, have asked the federal government for help: to grant them a waiver from the antitrust laws.
Since the federal government might not do that, or may be too slow to act, there is another solution that is already good law. It’s called Parker immunity. In Parker v. Brown (1943), the Supreme Court upheld a California program that limited the quantity of raisins producers could sell in order to stabilize prices. The Court held that a state “as sovereign” could, in essence, replace the Sherman Act’s mandate for competition with its own regulation. After all, “There is no suggestion of a purpose to restrain state action in the [Sherman] Act’s legislative history.”
The Court articulated a two-prong test for Parker immunity in California Liquor Dealers v. Midcal (1980). State regulations render private companies immune from federal antitrust liability if:
- The restriction is “clearly articulated and affirmatively expressed as state policy,” and
- It is “actively supervised” by the state.
This same framework could apply to AI safety agreements. A state could explicitly (through legislation) permit the AI labs to coordinate on uniform safety standards and restrict the release of models until they pass such tests. The statute should “clearly articulate[]” which companies and activities are covered and how this regime would work. Then, the state must “actively supervise[]” this policy, ensuring that it’s not just a vague “AI safety is important” statute. If the two conditions articulated in Midcal are met, the AI labs would be absolved of Sherman Act liability. And better yet, only one state need pass this. Just like most AI is “produced” in California, almost all of the country’s raisins were produced in California at the time of Parker, and over 90 percent were shipped out of state. Nonetheless, the Court upheld the regulation. As for which state it would have to be, since the leading AI companies are headquartered in California, the strongest argument for passing the Midcal test would be for California to pass it since it’s well-situated to actively supervise these companies. However, the law does not mandate that it be California. If any other state could make a convincing claim that they were actively supervising these companies, that regulation, too, could pass the test.
To be sure, this kind of policy could help incumbents entrench their market power. The leading AI labs could use safety as a pretense to lobby for regulations that would burden smaller companies and block their entry into the market, especially if legislation is not calibrated to specifically address the frontier. State officials must be careful to thoroughly and independently review proposed agreements and restrictions.
The importance of Parker immunity is underscored by the fact that a general plea for consumer safety may not be strong enough to insulate the AI companies. The Supreme Court confronted this issue head-on in National Society of Professional Engineers v. United States (1978), where an engineering association had banned its members from bidding against each other on price. The rationale was that allowing engineers to undercut each other would “produce inferior engineering work endangering the public safety.” After all, nobody wants buildings and bridges to fall down because of shoddy design. Nonetheless, the Court declined this argument: “Exceptions to the Sherman Act for potentially dangerous goods and services would be tantamount to a repeal of the statute.” So, too, in the AI context. Barring a repudiation of National Society of Professional Engineers by the Court, it’s unlikely that public safety arguments alone can win the day.
Parker is the better way. Nevertheless, this application of it remains untested. A state program may raise questions, especially as its reach crosses borders. Likewise, courts may find that a state’s supervision is inadequate, or that the state’s policy is not clearly articulated, assuming that the federal government or private plaintiffs choose to sue nonetheless.
One potential issue remains—even if a state program protects the AI labs from the Sherman Act, the Federal Trade Commission could separately take enforcement action against the labs for engaging in “unfair methods of competition” under Section 5 of the Federal Trade Commission (FTC) Act. The D.C. Circuit struck down an FTC attempt at overriding state optometry regulations, but the Supreme Court explicitly left open the question of whether Parker immunity applies to the FTC Act. So, such a program as laid out in this article could protect the AI labs from the Sherman Act but leaves the question of the FTC Act slightly less certain.
But the FTC’s own decisions provide substantial reassurance: it has dismissed a Section 5 complaint against a state-supervised private trade association on the basis of Parker immunity, and, in 2024, unanimously modified an order to permit a pharmacy cooperative’s joint negotiations because the conduct “qualifies for state action immunity to the extent that the entity appointed by Puerto Rico supervises the negotiations. . .” A carefully designed AI-safety program would therefore have substantial precedent supporting protection from FTC enforcement, although the Supreme Court has not conclusively resolved the outer limits of that protection.
Still, Parker immunity addresses the concerns of all those who believe the Sherman Act will block needed coordination. Under Parker, states can prioritize their own policy interests “as sovereign[s]” in lieu of federal antitrust laws. If no federal policy is forthcoming, Parker teaches us the answer is simple. A state—any state—must choose to act. Such action is certainly more readily forthcoming from the states, the laboratories of democracy, than Congress.
