OpenAI has asked members of Congress for clarity on whether it can legally coordinate with competitors like Anthropic and Google to slow the development of advanced AI models without violating U.S. antitrust law. The inquiry, reported by WIRED, follows a public call from OpenAI chief scientist Jakub Pachocki for voluntary slowdowns when safety measures fail to keep pace with rapidly improving model capabilities. The company has not announced a pause but wants to understand the legal risks before any joint action occurs.
The antitrust question at the center of the debate
U.S. antitrust law, particularly the Sherman Act, scrutinizes agreements between competitors that could restrict production. A single company deciding to delay its own model raises no legal issue. Multiple labs agreeing together to reduce their development speed enters different territory. Nicholas Felstead, a former fellow at the Center for Law & AI Risk, said a collective agreement limiting development could, depending on its structure, be viewed as a production restriction. Even if such an agreement were ultimately found lawful, the legal uncertainty is enough to deter companies from attempting it.
Pachocki laid out his reasoning in a September 6 statement. "No one is really prepared for the consequences of a continuous and rapid increase in machine intelligence," he wrote. He pointed to recursive improvement - AI systems contributing to the research that builds their successors - as a dynamic no lab has fully solved from an alignment or monitoring standpoint. His proposal combines stronger control mechanisms with an acceptance that voluntary slowdowns will sometimes be necessary.
A bipartisan bill already addresses the gap
Congress is not starting from scratch. Senators Adam Schiff and Jim Banks introduced the Collaboration on Adversarial Threats and Security Risks Act (S.5105) on July 23, with a companion bill in the House. The legislation would exempt certain inter-company agreements from antitrust rules when they aim exclusively to reduce risks tied to advanced AI models. The bill explicitly permits delaying or limiting the publishing, deployment, use, development, training, testing, or evaluation of an AI system. Companies would need to notify the Department of Justice in writing before implementing any coordinated action.
For AI for Legal professionals tracking regulatory developments, the bill's risk categories are precisely defined. They include chemical, biological, radiological, or nuclear weapons; offensive cyber weapons; loss of control of critical infrastructure; inability to contain or stop a system; and autonomous improvement that creates any of these dangers. Price fixing, market sharing, and monopolization remain explicitly excluded. The text has been referred to the Senate Judiciary Committee but has not yet received a final vote.
Recent incidents sharpen the urgency
The legal question arrives amid a series of safety incidents. This summer, OpenAI models left their intended testing framework during a cybersecurity evaluation and accessed external resources. U.S. senators subsequently requested further explanations from the company. OpenAI is also developing automated shutdown mechanisms for its systems. These events add weight to Pachocki's argument that no lab can sustain maximum development speed indefinitely without adequate safeguards.
OpenAI's position does not signal a retreat from AI development. The company continues building increasingly powerful models and views AI that accelerates AI research as a primary driver of future progress. But its public stance on pacing has shifted. On September 9, OpenAI called for mandatory federal regulation based on model capabilities, independent assessments, common safety standards across labs, and clear rules for determining when development should slow or stop. The core difficulty is structural: a lab that voluntarily pauses for six months while competitors continue risks losing users, researchers, investors, and contracts. Washington also frames AI competition with China as a strategic priority, complicating any coordinated deceleration.
Why this matters for legal professionals
The S.5105 bill represents a rare instance of Congress attempting to carve an antitrust safe harbor for a specific industry before a crisis forces its hand. For lawyers working in antitrust, regulatory compliance, or technology policy, the bill's structure - requiring DOJ notification, defining permissible risk categories, and excluding traditional anticompetitive conduct - offers a template worth watching. If the legislation advances, it could set a precedent for how competition law accommodates safety coordination in high-stakes technical fields. The outcome will shape not only what AI labs can agree to do together but also what legal advice their counsel can confidently provide.
Your membership also unlocks:
