Following reports of AI agent swarms hacking websites and communicating via secret message boards, leading AI companies have proposed a coordinated slowdown in AI development. They have expressed concerns that this action could conflict with antitrust laws. Antitrust experts suggest that while the companies' language may not be beneficial, the uncontrolled development of advanced AI may not align with the Sherman Act, a key US antitrust law designed to foster competition. Obtaining government approval for continued development could help avoid costly investigations in the future.
Under antitrust law, how employees discuss business decisions can be as significant as the decisions themselves. Google has previously trained its employees to avoid phrases that could imply anticompetitive behavior, instead encouraging them to focus on how business decisions improve offerings for consumers.
Using terms like "slowdown" or "pause" may raise more concerns than the actual activities involved in ensuring advanced AI models remain safe. A collective agreement to slow down without a clear purpose could be interpreted by regulators as an anticompetitive agreement. John Bergmayer, legal counsel for Public Knowledge, noted that the companies may have limited their options by their choice of language.
Bergmayer explained that antitrust economists often assess whether companies are reducing output, suggesting that AI firms could have framed their intentions around safety protocols instead of implying a slowdown.
Meta CEO Mark Zuckerberg did not explicitly endorse the slowdown but stated that AI labs have a strong incentive to ensure AI agents behave properly, as consumer demand is for models that align with user intentions. He emphasized that companies that fail to ensure alignment may fall behind in the market.
David Lawrence, a former policy director at the Department of Justice’s Antitrust Division, mentioned on LinkedIn that agreements aimed at preventing catastrophic risks could actually promote competition and are protected under the "ancillary restraints doctrine."
Concerns about collective agreements not to implement safety measures could lead to allegations of "quality fixing," as highlighted by Roger Alford, a professor at Notre Dame Law School. He referenced a European antitrust case where car manufacturers collaborated on emissions technology but agreed not to compete on improvements beyond legal requirements, resulting in significant fines.
Bergmayer noted that industries can limit antitrust liability through the National Cooperative Research and Production Act of 1993, which allows them to establish standards-development organizations with notification to the FTC and DOJ.
Skepticism exists regarding antitrust exemptions, as they may benefit larger companies and hinder new entrants. David Sacks, cochair of the President’s Council of Advisors on Science & Technology, criticized the request for an exemption, labeling it a strategy to form a cartel.
Many employees at major AI companies have expressed concerns about the rapid pace of model development and its potential safety implications. Bergmayer commented that companies sometimes seek regulation as a strategy to maintain their market position, while others genuinely feel pressured to act against their better judgment.
Both Anthropic and OpenAI have filed confidential paperwork for initial public offerings, with valuations nearing or exceeding one trillion dollars. Anthropic is expected to go public next month, while OpenAI plans to delay its IPO until 2027 due to safety concerns. The Ramp AI Index indicates that both companies are competing closely in terms of model usage.
Alford remarked that the companies may seek an exemption to coordinate a slowdown, fearing that unilateral action could disadvantage them in a rapidly advancing market. Political pressure also exists to accelerate AI development, as indicated by President Donald Trump's comments on social media regarding the government's regulatory power over AI companies.
The DOJ Antitrust Division faces its own political challenges, with Alford having been ousted after criticizing the approval of a merger he deemed corrupt. Reports suggest that Andrew Ferguson, chair of the FTC, indicated that AI policy decisions would ultimately be influenced by Trump.
If the federal government were to investigate an AI slowdown, it would be classified as a "conduct investigation," which can take years and involve extensive document production and employee depositions. As no new regulations are anticipated and an antitrust exemption seems unlikely, AI companies will need to establish their own guidelines while navigating potential investigations.