Recent reporting describes employees at competing AI companies talking and strategizing across company lines about how to use their positions to steer the industry in a safer direction. Some employees may worry that these conversations create antitrust exposure, particularly in light of a lawsuit filed against their employers alleging a proposed agreement to slow the pace of AI development.
Generally speaking, it is perfectly fine for employees at competing companies to communicate with each other about subject matter relevant to their work. But they should not share competitively sensitive information like non-public prices, the scale of planned training runs, model-release timing, unreleased product features, or compute capacity. And conversations among employees cannot serve to facilitate collusion between the companies.
There are two main ways in which communications between employees of competing companies can pose a problem under the antitrust laws. “Communications” here is broad — I am referring to conferences, emails, group chats, discussion groups, happy hours — any communications between employees of competitors that bear on their companies and their work. First, such communications can be used as circumstantial evidence in support of a conspiracy between the companies to fix prices or reduce output, which is a per se violation of Section 1 of the Sherman Act. Second, information-sharing between competitors itself can be an illegal restraint of trade in violation of Section 1.
I cannot analyze the risk level of every conceivable subject of discussion, though I identify some especially common categories at the end of this post. To build an intuition for which conversations are benign and which are risky, I describe below how inter-company employee communications could be offensively wielded to support each theory.
For background, I am a lawyer. I practiced at a firm for five years where I did a lot of antitrust law. This document is not meant as legal advice.
Communications used to support the existence of a conspiracy
To establish a conspiracy to fix prices or reduce output in violation of Section 1, a plaintiff would need to show that the companies took parallel action and present some evidence implying a preceding agreement. So, the scenario in which the inter-company communications are relevant for purposes of this legal theory is one in which the AI companies have taken parallel, output-restricting action, and the plaintiff seeks to prove that the companies secretly agreed with each other to do so beforehand. (If the companies independently chose to do so, that would be legal. If they agreed with each other to do so, that would be illegal.) For example, Dario Amodei announces that Anthropic is going to pause training runs above a certain threshold. Days later, Sam Altman announces that OpenAI will do the same. A plaintiff might investigate meetings held, emails sent, and texts exchanged between employees at the two companies in the lead-up to their announcements for evidence of an agreement.
Employees should therefore be careful not to act as intermediaries for the companies. That means, for example, that company A’s employees should not tell company B’s employees that company A will take an action if company B does, even if — especially if — it is true. Such communications could be used to prove a Section 1 conspiracy if the companies subsequently take the action.
The mere fact of communications between line engineers and other typical employees, by itself, however, is perfectly innocuous1. Courts understand that inter-company employee communications are common and often pro-competitive, as when researchers share new relevant research that leads to product improvements. Of course, this would not necessarily hold for communications between C-suite executives and senior management. A record of unexplained phone calls between Dario and Sam before they each announced a pause would look very suspicious.
Suppose employees at the AI companies got together and agreed to urge their employers to adopt certain safety practices or to pause training, and the companies agreed in order to appease their employees. Would that constitute an illegal conspiracy between the companies? Without more, no. Section 1 distinguishes between parallel conduct in response to common market stimuli and a conspiracy2. While I’m not aware of any antitrust cases involving competitors taking parallel actions in response to their employees’ parallel demands, cases make clear that competitors making parallel pricing or output decisions in response to common market forces do not thereby violate Section 1. For example, in In re Musical Instruments & Equipment Antitrust Litigation,3 the court held that the plaintiffs failed to show a conspiracy when a powerful retailer pressured competing guitar manufacturers to adopt the same pricing policy and each independently acquiesced. “The manufacturers’ similar response to this market pressure is a hallmark of independent parallel conduct—not collusion.”
Information-sharing between competitors
Information-sharing itself may also violate the antitrust laws, but it is not a per se violation like a conspiracy to restrict output. Whether it is illegal depends on the type of information that’s shared. Keep in mind that the reason information-sharing is suspect under the antitrust laws is that it may facilitate collusion between companies. So sharing confidential information about pricing, future training plans, model release timing, features of unreleased models, key inputs like compute capacity, and so on is very risky.4 (In addition to antitrust concerns, this may also be a trade secret violation.) But discussing and sharing information like assessments of one’s employer, policy ideas, the news, interpretability methods, and fundamental research findings — these are all fine.5 This type of information does not obviously help the companies coordinate to increase prices or reduce output.
Employees’ own organizing activity wouldn’t violate the antitrust laws
One concern AI company employees might have is that by coordinating to refuse certain types of work or to quit altogether in protest of their companies’ actions, they are personally violating the antitrust laws. After all, it’s an agreement among competitors (employees) to withhold output (labor). But this is not an antitrust problem. In Allied International, Inc. v. International Longshoremen’s Association,6 stevedores jointly refused to handle Soviet cargo as a protest against the Soviet invasion of Afghanistan. The court found that their activities did not constitute an antitrust violation. Their group refusal to handle goods was “undertaken as a political protest” and “ill-designed as a means of gaining a competitive or commercial advantage for the union or its members,” and thus was “not the sort of evil at which the Sherman Act is aimed.”7 Of course, nothing about this precludes the AI companies from firing employees who engage in such coordinated action.
Practical upshot
Discussions between employees of competing AI companies generally pose little antitrust concern. Here are things employees can do:
Discuss AI safety
Share policy ideas and research
Discuss their opinions about their employers
Coordinate advocacy aimed at persuading their respective companies to adopt safer practices
Discuss quitting, speaking out, or other actions they should take to pressure their company leadership
There are two main things to avoid. First, employees should not transmit confidential or competitively sensitive information about their company’s commercial plans: things like pricing, the scale of planned training runs, model-release timing, unreleased product features, or compute capacity. Second, employees should not act as intermediaries through which their companies can coordinate their conduct. In particular, employees should not communicate commitments, assurances, or conditional offers along the lines of “our company will do X if yours does too.”
A useful rule of thumb is: employees can coordinate with each other about what they want their companies to do; they should not coordinate the companies themselves. The closer a conversation gets to exchanging confidential plans or transmitting commitments between company decision-makers, the greater the antitrust risk.
See, e.g., Krehl v. Baskin Robbins Ice Cream Co., 664 F.2d 1348, 1357 (9th Cir. 1982) (finding “[e]vidence of sporadic exchanges of shop talk among field sales representatives who lack pricing authority” insufficient to support a price-fixing conspiracy).
Bell Atl. Corp. v. Twombly, 550 U.S. 544, 556 n.4 (2007).
798 F.3d 1186, 1195 (9th Cir. 2015).
See, e.g., Todd v. Exxon Corp., 275 F.3d 191, 211 (2d Cir. 2001) (citing American Column & Lumber Co. v. United States, 257 U.S. 377, 398–99 (1921)).
See Maple Flooring Mfrs.’ Ass’n v. United States, 268 U.S. 563, 586 (1925).
640 F.2d 1368 (1st Cir. 1981), aff’d, Int’l Longshoremen’s Ass’n v. Allied Int’l, 456 U.S. 212 (1982).
Id. at 1380.