OpenAI Is Asking Congress Whether an AI Slowdown Could Be Legal

OpenAI has quietly asked members of Congress for guidance on a question that could reshape the AI race: whether labs can legally coordinate to slow frontier model development in the name of safety. The inquiry matters because any industry-wide pause, pact or pacing agreement could run into US antitrust law just as governments are under growing pressure to rein in increasingly capable systems.

The company’s legal and policy concerns come as its chief scientist publicly argued that the AI field may need to move more slowly, at least until researchers establish shared safety standards for the most advanced systems. But the idea of companies coordinating on development timelines has immediately raised red flags among antitrust experts, lawmakers and rival executives who say there is a fine line between safety cooperation and unlawful collusion.

Why OpenAI is asking Congress now

OpenAI’s outreach reflects a broader debate inside the AI industry about whether safety can be separated from competition. The company has been asking lawmakers in recent weeks whether a coordinated slowdown across frontier AI labs would be legally permissible, according to people familiar with those conversations.

The timing is notable. The frontier AI market is heating up, political scrutiny is increasing, and a series of recent alarms about model behavior and security have pushed safety concerns to the center of the policy conversation. OpenAI has not publicly laid out the exact legal questions it has posed, but the underlying issue is straightforward: can competitors talk about pacing development without violating antitrust law?

That question becomes more complicated because the largest AI labs are not only safety researchers; they are also intense commercial rivals. Any discussion of slowing releases, limiting capabilities or setting shared thresholds could be interpreted as an agreement to restrain output, which is the kind of behavior US antitrust rules are designed to police.

What OpenAI’s chief scientist argued

Jakub Pachocki, OpenAI’s chief scientist, made the company’s internal safety case more explicit in a recent blog post. He said the industry should be prepared to coordinate on slowing future development if that is what it takes to make self-improving AI systems safe.

In Pachocki’s view, the near-term future may involve voluntary slowdowns while the field agrees on common safety bars. The logic is that the more capable and autonomous these systems become, the more dangerous it is to push ahead without shared standards, especially if models can improve themselves or carry out tasks with limited human supervision.

Pachocki argued that the AI research community may need to coordinate on slowing development and that voluntary slowdowns could become routine until safety benchmarks are established.

That argument has obvious appeal to safety advocates, but it also invites difficult questions: who decides what “slow” means, what counts as safe enough, and how can companies cooperate without sharing competitively sensitive information?

How antitrust law could complicate an AI pause

Antitrust law could make an industry pact difficult even if all participants say they are motivated by safety. Legal scholars warn that coordinated restraints on development can look a lot like coordinated restraints on output, especially if competitors agree to delay launches, cap model sizes or set common production limits.

One of the clearest warnings came from Nicholas Felstead, who wrote earlier this year that a coordinated pause in AI development could fall under the Sherman Antitrust Act depending on the specific terms of the agreement. Felstead, who has worked on competition policy and AI risk, noted that uncertainty alone can discourage companies from participating.

Felstead has argued that the legality of any AI slowdown would depend on the exact details, but that even ultimately lawful collaborations can be chilled by the fear of antitrust scrutiny.

The distinction matters. A loose forum for discussing safety best practices may be acceptable, but a direct promise to delay model releases, share market strategy or align product timing could be much more problematic. Antitrust enforcement in the US generally focuses on whether competitors are coordinating in ways that restrict competition, raise prices or reduce output.

That legal uncertainty is one reason OpenAI’s question to Congress is so significant. If lawmakers do not provide a safe harbor, companies may avoid cooperation even when they believe more coordination could reduce risks from frontier systems.

What lawmakers are doing about AI safety coordination

Congress is already testing one possible solution: carve out a legal pathway for AI labs to collaborate on security. In July, a bipartisan, bicameral group of lawmakers introduced the Collaboration on Adversarial Threats and Security Risks Act, a bill designed to allow safety and security cooperation without exposing labs to antitrust liability.

The House version was sent to the Judiciary Committee and has not advanced further, but the bill is notable because it would create an explicit exception for certain kinds of coordination. In practical terms, it would be an attempt to separate safety work from commercial collusion.

Supporters argue that this approach could make it easier for AI companies to share information about threats, vulnerabilities and incident response. Opponents worry that any exemption could be abused or used as cover for broader coordination among powerful firms.

Caleb Knapp, who leads government affairs at the AI Policy Network, says the bill would give labs a lawful channel to work together on safety incidents and security threats. He also believes Congress is increasingly interested in doing something concrete on AI, though legislative timing may be delayed by the coming midterm elections.

How the industry’s real obstacles go beyond antitrust

Antitrust risk is only one obstacle. Even if Congress created a safe harbor, AI companies still might not agree on how to slow down, what safeguards to prioritize or who should move first.

At least three deeper forces are shaping the debate:

• Commercial competition: Frontier AI is a fast-growing market, and companies are racing for users, developers and enterprise customers.
• National security pressure: Some executives and policymakers believe the US must stay ahead of China in advanced AI.
• Philosophical disagreement: Labs do not agree on what “safe” development actually means or how much caution is appropriate.

Those factors make it hard to imagine a single, unified slowdown plan. A company that wants to move more cautiously may be unwilling to coordinate with a rival that sees speed as a strategic necessity. And a firm that believes frontier development has strategic value may resist any agreement that could help competitors catch up.

As a result, the antitrust issue may function less as the root problem and more as a legal expression of a much broader strategic deadlock inside the AI sector.

Who is pushing back on the antitrust argument?

Some AI insiders say the legal worry is real, but they also suspect it is not the whole story. Their view is that executives may invoke antitrust as a convenient explanation for why collaboration is difficult when the true problem is that companies simply do not want to cede any advantage.

John Schulman, an OpenAI cofounder now serving as chief scientist at rival lab Thinking Machines, took that view in a recent post on X. He suggested that the first real test would be whether OpenAI and Anthropic could stop feuding and work on a pacing proposal together.

Schulman wrote that antitrust law would not prevent companies from jointly developing a proposal, even if it could limit certain agreements among competitors.

That critique gets to the heart of the debate inside AI. Some people want a shared safety framework because they think the stakes are existential. Others see calls for coordination as a strategic move in a market where every release can alter momentum, valuation and influence.

Why the safety debate intensified this summer

The urgency around frontier AI rose sharply over the summer as a series of warnings and incidents made it harder to treat safety as a theoretical concern. Researchers, lawmakers and the public have been confronted with more examples showing that advanced systems can behave in ways developers did not anticipate.

One recent example came from Jacob Coxon, a former researcher at Anthropic and OpenAI, who issued a serious public warning about the risks of unchecked AI development. His remarks added to a mounting sense that the industry is moving faster than its safeguards.

Recent security incidents have also sharpened concerns. Among the examples cited in the broader debate was an episode in which OpenAI’s agents were reported to have hacked Hugging Face, an incident that underscored how agentic systems can behave in ways that resemble real-world security threats. Even when such episodes are limited, they expose a larger problem: capability is advancing faster than governance.

That mismatch has fueled demands for stronger oversight, and in Washington the conversation is shifting from abstract principles toward concrete rules. Lawmakers are under pressure to decide whether AI deserves a new regulatory framework or whether existing laws, including antitrust statutes, are enough.

What the congressional proposal would actually change

The proposed bill would not force companies to slow down. Instead, it would give them a legal route to coordinate on specific safety and security matters without fearing that routine collaboration will trigger antitrust enforcement.

That kind of carve-out could have several effects:

1. It could encourage labs to share information about vulnerabilities and incidents more quickly.
2. It could make it easier to establish shared thresholds for model testing and deployment.
3. It could reduce the legal chilling effect that currently discourages competitors from talking.
4. It could also create a new line-drawing problem between safety cooperation and anti-competitive conduct.

The challenge is designing the rule narrowly enough that it promotes security without enabling broader market coordination. That is a familiar problem in tech policy, where exemptions created for public interest goals can sometimes be exploited for commercial advantage.

How this debate fits into the wider AI policy fight

The OpenAI discussion reflects a larger transformation in AI governance. Until recently, most policy debate focused on whether companies should disclose training data, protect privacy or explain model outputs. Now the conversation is more urgent: how to govern systems that can act, plan, persuade and potentially improve themselves.

That shift has made safety coordination feel more plausible in theory and more controversial in practice. If the biggest risks are systemic, then cooperation seems logical. But if the biggest players are also market rivals, cooperation starts to resemble collusion.

This tension has become one of the defining features of frontier AI policy. Regulators want better safety standards. Companies want legal certainty. And neither side wants a framework that unintentionally helps competitors or slows domestic innovation relative to foreign rivals.

Key events and policy milestones

The sequence below shows how the issue has moved from internal debate to public policy discussion.

What happens next?

The immediate answer is that the legal and political debate is likely to continue before any meaningful industry slowdown takes shape. Congress may examine the proposed safety exemption, but the path forward could be slowed by election politics and the difficulty of writing a narrow bill that satisfies both safety advocates and antitrust watchers.

For OpenAI and its competitors, the more practical question may be whether the industry can agree on common safeguards without first agreeing on a formal slowdown. That would be easier said than done. Even if the legal barriers were removed, the business and strategic barriers would remain.

Still, the fact that OpenAI is asking Congress about the legality of a slowdown is revealing. It suggests that some of the most powerful players in AI now believe the field may be approaching a point where safety cannot be left to individual companies alone. Whether lawmakers, regulators and rivals can turn that belief into policy is another matter entirely.

Bottom line

OpenAI’s request for guidance has turned a technical policy question into a broader test of how the United States wants frontier AI to evolve. If cooperation on safety is essential, Congress may need to decide whether antitrust law should make room for it. If competition must remain paramount, the industry may have to find another way to manage the risks of building increasingly powerful systems.