Concerns Over AI Development and Coordination
In light of recent reports regarding swarms of AI agents hacking websites and collaborating in secret forums, coupled with warnings from a former engineer at Anthropic, leading AI firms have urged for a coordinated “slowdown” in the advancement of this technology. While various companies have articulated this notion in differing ways, they have also voiced apprehensions that such an agreement might breach antitrust laws.
Competition experts argue that although the manner in which these companies have framed the discussion may not work in their favour, the unrestricted development of potentially uncontrollable AI systems does not align with the objectives of the Sherman Act—one of the principal antitrust laws in the United States designed to safeguard competition. Simultaneously, obtaining formal approval from the government could help circumvent costly investigations in the future.
Antitrust Implications of Language Used
In the realm of antitrust, the terminology used by employees to describe business decisions can be almost as significant as the decisions themselves. For instance, Google has trained its staff to avoid phrases that could imply anti-competitive practices, even in internal communications. Instead, employees are encouraged to emphasise how these choices enhance their products and benefit consumers.
From an antitrust perspective, phrases like “a slowdown” or “a pause” may raise more suspicion than the actual activities they reference, such as developing safeguards to prevent advanced AI models from going awry. A collectively agreed slowdown without a clear objective might be interpreted by regulators as an anti-competitive agreement aimed at reducing trade.
As noted by John Bergmayer, legal advisor for the non-profit organisation Public Knowledge, “In a way, they have painted themselves into a corner with the way they have phrased things.” He elaborates that economists often examine whether production is being curtailed, implying that if two or more companies are coming to an agreement to “take it easy,” it could signal anti-competitive behaviour. Rather than framing the initiative as a means to curtail development, AI firms could have emphasised their collaboration in establishing safety protocols against catastrophic risks. Consequently, any slowdown in the release of new models would appear as a by-product of these measures, rather than the primary aim.
Divided Opinions Among Industry Leaders
Mark Zuckerberg, CEO of Meta, recently addressed the proposal for a slowdown without endorsing an explicit “slowdown.” Instead, he advocated that AI laboratories possess a strong natural incentive to ensure their agents are well-behaved, as consumers prefer models that do not act unpredictably—a situation referred to in AI terms as “misalignment.” He asserted that companies failing to invest the necessary time to achieve proper alignment “will fall behind” competitively. This is akin to two car manufacturers declaring, “we have agreed not to make better cars for a while,” compared to saying, “we will not produce faster vehicles until we can ensure they are safe, as no one will buy them if they are dangerous.”
David Lawrence, who recently served as the director of policy for the Antitrust Division of the Department of Justice (DOJ), commented on LinkedIn that agreements designed to avert catastrophic risks “actually increase production and foster competition,” and are already protected under what is known as the “doctrine of ancillary restraints.” A FTC antitrust lawyer concurred, stating, “After all, without competition, there would be no humanity.”
Legal Risks of Collective Safety Measures
However, a collective agreement not to implement safety measures could expose AI laboratories to allegations of “quality manipulation,” as noted by Roger Alford, a professor at Notre Dame Law School and former deputy director of the Antitrust Division at the DOJ. This occurs when companies mutually agree not to enhance their products. Alford referenced a European antitrust case where automotive firms collaborated to develop emission-reduction technology while agreeing not to compete on improvements beyond legal requirements, ultimately leading to a fine of about one billion dollars.
Self-regulation is not a novel concept. Bergmayer points out that industries can already mitigate their antitrust liability under the National Cooperative Research and Production Act of 1993, which allows them to form a standards development organisation as long as they notify the FTC and DOJ.
Market Pressures and Regulatory Skepticism
Advocates for competition often express scepticism towards antitrust exemptions, claiming they tend to empower larger players and inhibit newer companies from gaining market traction. Others, including David Sacks, co-chair of the Presidential Advisory Council on Science and Technology, have also questioned the request, accusing Anthropic and OpenAI of forming a duopoly and describing the antitrust exemption request as a “psychological operation in a full-blown electoral campaign” and an excuse to “form a cartel.”
Nonetheless, many employees within leading AI companies have voiced concerns regarding the rapid pace at which new models are being produced, potentially at the expense of safety. “Sometimes companies seek regulation as a strategy to protect themselves. Yet, they often feel that the market pressures them to act in ways they deem inappropriate,” Bergmayer observes. He adds, “Psychologically, perhaps AI labs feel they lack support from the markets, their upcoming IPO, or other circumstances that would allow them to take unilateral actions.”
Impending Initial Public Offerings and Market Competition
This summer, both Anthropic and OpenAI submitted confidential documentation for their initial public offerings, each reaching a valuation near or exceeding one trillion dollars. Anthropic is expected to go public next month, while OpenAI’s CEO, Sam Altman, announced plans to postpone their IPO until 2027 due to recent safety concerns. Reports suggest that OpenAI’s CFO has already informed employees that the company would list publicly in 2027. The Ramp AI Index, which tracks AI adoption rates, indicates that both companies have similar usage rates and are fiercely competing for the top position. Neither company responded to requests for comment regarding this article.
Alford notes, “I think they want an exemption because they believe they need to coordinate to slow the pace down, as they do not want to disarm unilaterally while others continue to advance at breakneck speed.” Moreover, there exists political pressure to progress as swiftly as possible. Following the news of the proposed slowdown, former President Donald Trump stated on social media that the government possesses “enormous PENAL and REGULATORY power over these companies!” and that “WHOEVER WINS IN AI, WINS!”
Regulatory Investigations and the Future of AI Development
Additionally, the Chief Technology Officer of the Department of Defence has been posting “anti-catastrophist” memes in apparent response to the notion of a slowdown. The Antitrust Division of the DOJ is also not immune to political pressures. Alford was ousted from the agency last year after accusing DOJ leadership of corruption for approving a merger between two tech infrastructure companies, an action he later referred to as a “scandal” in a public speech.
Moreover, a journalist from The Wall Street Journal reported that FTC Chair Andrew Ferguson stated during a competition conference that Trump “would decide any policy regarding AI.” If the federal government were to initiate an antitrust investigation concerning an AI slowdown, it would focus on potential anti-competitive practices. Unlike merger reviews, which are subject to strict timelines, such investigations could extend over years. AI companies might be compelled to produce millions of pages of documents, while executives and other key employees could be required to testify. The devices of hundreds of workers might also be subject to legal preservation, even if the government ultimately concludes that no charges are warranted.
With no new regulations on the horizon and an antitrust exemption seeming unlikely, pioneering AI laboratories will need to draft their own operational guidelines while potentially navigating various investigations along the way.