OpenAI Asks Congress About Legality of Slowing AI Development

It has come to light that OpenAI had been seeking clear guidance from the U.S. Congress on the legality of coordinating an industry-wide slowdown in frontier AI development. According to reporting by Wired’s Maxwell Zeff, citing people close to the company, the outreach over the past few weeks was detailed. The background concern is that even coordination aimed at safety could run afoul of antitrust law.

At the heart of the issue is whether an agreement among competing AI labs on development pace or release timing would amount to a restriction of output. In the United States, the Sherman Antitrust Act prohibits output restrictions and market allocation among competitors. Even when done in the name of safety, legal risk remains unless an exemption is explicitly codified.

OpenAI Chief Scientist Jakub Pachocki, in a blog post published last weekend, cited coordinated slowdown as the direction the research community should take. He argued that pacing the development race is essential to keeping self-improving AI systems safe. In the short term, he said voluntary slowdowns would become the norm until common safety standards are established.

coordinating to slow down future development

The above is a central passage from Pachocki’s post. It envisions a phase before shared safety benchmarks are in place, in which each company slows its pace on its own judgment. Conversely, it also points to the need to advance standard-setting and pace coordination across the industry.

In response to the proposal, legal scholars have warned of a collision with antitrust law. Nicholas Felstead, an assistant director at the Australian Competition and Consumer Commission and a former AI policy fellow at the Center for Law & AI Risk, argued in a March essay that a coordinated development pause could qualify as an output restriction. Depending on the details of the agreement, it could implicate the Sherman Antitrust Act, he analyzed.

Felstead wrote that even if much safety coordination would ultimately survive scrutiny, the uncertainty itself acts as a powerful deterrent. When legality is unclear in advance, corporate legal departments become cautious about joining coordination. He noted that this chilling effect is an obstacle to involving giant tech companies. OpenAI reportedly did not respond to requests for comment before publication.

Rise of the Frontier AI Slowdown Debate and

Its Background

Capability gains in frontier models have entered a new stage, driven by improved reasoning performance and expanded autonomous functions. There is concern that once models begin to run their own improvement loops, traditional ex post evaluations will not detect dangers in time. Pachocki’s warning aims to address this time lag.

The current state of the development race is also part of the background. Training foundation models requires massive computing resources and talent, creating a structure with large first-mover advantages. Companies are competing to update flagship models, as seen in OpenAI Announces GPT-5.6 with Three Models: Sol, Terra, and Luna, and are locking in the application layer. Fear of being overtaken by rivals if they slow down makes coordination difficult.

The reasons shared safety standards have not advanced are also complex. Labs differ on evaluation methods, definitions of dangerous capabilities, and pre-release verification levels. Without agreement on what counts as safe, the conditions for slowing down cannot be set. Technical benchmarking and legal backing are needed at the same time.

The pace of product rollouts also raises the difficulty of coordination. OpenAI continues to review its applied products, as in OpenAI to Shut Down AI Browser Atlas, Moving Features to Chrome Extension and App, with foundation development and product launches linked. Slowing development would affect not only research schedules but entire business plans. The weight as a management decision is growing.

How Antitrust Law Blocks Safety Coordination

U.S. antitrust law strictly prohibits competitors from agreeing in advance on prices, output, or sales territories. An industry-wide halt or speed limit on AI development could be interpreted as suppressing the supply of new models. Even if the purpose is safety, it would be subject to review without an explicit exemption in the text of the law.

Differences depending on the content of coordination are also important. Defensive information-sharing, such as sharing incident information or notifying vulnerabilities, is generally considered a more permissible area. By contrast, agreements to align training start dates, release dates, or levels of computing investment could be seen as close to output coordination. The ambiguity of the line dulls corporate decision-making.

Estimates of legal risk also hamper coordination. Antitrust violations carry heavy fines, civil lawsuits, and potential liability for executives. Even if illegality is later rejected, the burden of responding to investigations remains. Without a reliable safe harbor, the larger the company, the more cautious it will be about participating.

This structure is amplified by circumstances specific to the AI industry. A small number of labs lead frontier development, and their moves directly affect the entire market in an oligopolistic structure. Because an agreement among a few parties can easily affect supply across the whole market, antitrust authorities take a strict view. The distinction between well-intentioned safety coordination and anticompetitive restraint is at issue.

Aims and Limits of the Bipartisan CATS Act

Amid this stalemate, a bipartisan, bicameral group of lawmakers introduced in July a bill titled the “Collaboration on Adversarial Threats and Security Risks Act.” It would make clear that coordination among AI labs to address adversarial threats and national security risks can be carried out without fear of antitrust violations. The House bill was referred to the Judiciary Committee but has not proceeded to deliberation.

The bill aims to provide a legal pathway for collaboration involving security and safety. The nonprofit AI Policy Network expressed support for the bill. Caleb Knapp, the group’s head of government affairs, explained the significance of establishing a framework for collaboration to address safety and national security cases. The aim is to advance incident response and threat-information sharing without a chilling effect.

Knapp said momentum for AI safety efforts is growing within Congress. At the same time, he noted that legislation is likely to come only after the midterm elections. The reality that the political calendar dictates the speed of legislation is coming into view. In pre-election sessions, competition for priority is intense, and highly technical bills tend to be pushed back.

Even if the bill passes, defining its scope will remain a challenge. While allowing coordination on security and incident response, extending it to development pace or market-launch timing would heighten tension with competition policy. The drafting of how far the safe harbor extends will be key. Consistency with operating guidelines and interpretations by the Department of Justice and the Federal Trade Commission will also be needed.

The Clash Created by Competition and

National Security

There is also a persistent view that motives beyond antitrust law underlie caution toward coordination. Another camp of AI executives argues that legal concerns are being used as a pretext to avoid coordination. They contend the real barriers are business competition, national security, and divergent views on safety.

The first factor is the size of the market. Foundation models and their applications are a huge market in its early stages, and companies are competing to secure share. Curbing development speed means deferring revenue opportunities. For listed companies and heavily funded startups, accountability to shareholders and investors is also heavy.

The second factor is the perception of competition with China. Within the administration, there is a view that maintaining advantage in AI is central to national security. Some executives share this view. Concern that slowing down would help rival nations catch up strengthens resistance to coordination. Reconciling technological dominance with safety is a difficult challenge.

The third factor is differences in views on safety. Labs are deeply divided over the safe development path. Some argue for thicker pre-release verification, while others emphasize improvement through real-world use. There is also no settled definition of what counts as dangerous. Building a common benchmark itself becomes a political negotiation.

Editorial Opinion

In the short term, we expect deliberation on the CATS Act to remain stalled. We assess that contention in the Judiciary Committee is likely to slip past the midterm elections. We see little room for a standalone vote on AI safety in the pre-election session. We assess that companies will hold off on full-scale coordination, citing the absence of legal guidance.

In the long term, we see institutionalization of safety standards changing the conditions of competition. We assess that once common evaluation methods and pre-release verification are established, agreements to slow down will no longer be needed. We see compliance with standards becoming a precondition for market entry, shifting competition from speed to quality. We assess that transparency of incident response will improve for users.

The remaining question is who decides the scope of coordination. Should pace coordination be included in the safe harbor. How should advantage over China be reconciled with shared human safety. Can voluntary slowdowns left to labs be effective. We assess that designing the intersection of law, technology, and strategy is now at issue.

References

• “OpenAI Wants to Know if an AI Industry Slowdown Would Even Be Legal”, by Maxwell Zeff — Wired, 2026-09-10T23:28:42.000Z (ARR)
• Source URL: https://www.wired.com/story/openai-wants-to-know-if-an-ai-industry-slowdown-would-even-be-legal/