OpenAI Asks Congress if a Joint AI Slowdown Would Be Legal

OpenAI has asked lawmakers whether an agreement among AI companies to slow development together would violate antitrust law, according to Wired's reporting, cited separately by Decrypt on Sept. 11 and by TechRepublic on Sept. 14. Both outlets said the outreach followed a Sept. 6 essay by OpenAI chief scientist Jakub Pachocki calling for the industry to consider coordinated restraint.

What Pachocki's essay actually argues

In the essay, titled "An Alien Mind", Pachocki wrote that "our evaluations indicate our ability to rely on" chain-of-thought monitoring, the practice of reading a model's step-by-step reasoning to catch unsafe behavior before it acts, "is progressively diminishing." He wrote that "if AI progress continues, machine recursive self-improvement (RSI) will be at the very core of future scientific discovery," and said he "expects and hopes for voluntary slowdowns to become commonplace until shared safety bars are established." Pachocki also wrote that "international coordination on future AI development needs to become a top priority for governments around the world," naming a network of third-party auditors, government agencies or international bodies as possible enforcers.

The legal question underneath a safety essay

One company slowing its own release schedule raises no antitrust issue. Several competitors agreeing to do it together is a different matter: "under the Sherman Antitrust Act, formal agreements between competing corporations to throttle product rollouts can be interpreted as illegal output restrictions," TechRepublic wrote, describing the 1890 law OpenAI is asking Congress about. The law bars competitors from jointly restricting output or dividing markets no matter the motive behind it, which is why a voluntary pact among OpenAI, Anthropic and Google to pace their own releases could draw the same scrutiny as a price-fixing agreement even if every participant calls it a safety measure.

Miranda Bogen, chief technologist at the Center for Democracy and Technology, a Washington nonprofit focused on digital rights, told Decrypt that "commercial and geopolitical competition in the AI space is incredibly intense, leading to a concerning dynamic where companies are incentivized to release products before their risks are fully understood." She said that "even when internal staff knows more research and testing is needed, their companies are facing immense pressure to cut corners and skip critical safety tests, despite evidence piling up about the consequences of moving too fast."

Duncan Sabien, head of communications at the Machine Intelligence Research Institute, a Berkeley, California, nonprofit focused on AI safety research, told Decrypt that "sans some sort of coordination mechanism, stepping back just means the other guy gets a lead," adding that "gains in intelligence are compounding; each new system makes it easier to train and deploy the next, more-advanced system." Sabien said the underlying incentive is financial as much as competitive: "every advance under current conditions yields many millions or billions more in funding and puts the creators of that advance in a greater position of power and influence." That argument mirrors the one at the center of Y Combinator president Garry Tan's push to let U.S. labs distill frontier models: a single company holding back rarely changes the industry's pace, only its own position in it.

A narrower bill already sits in Congress

Senators Adam Schiff and Jim Banks introduced S.5105, the Collaboration on Adversarial Threats and Security Risks Act, on July 23, 2026, according to GovInfo's bill text. The bill creates a narrow antitrust safe harbor, but only for five defined risk categories, and only if companies notify the Justice Department in writing first.

S.5105 safe-harbor category	What it covers
Weapons of mass destruction	Chemical, biological, radiological or nuclear risks
Offensive cyber weapons	Coordinated defense against cyberattack capability
Critical infrastructure	Loss of control of infrastructure systems
Containment failure	Inability to contain or stop a system
Autonomous improvement	Self-improvement that creates any of the above

Price-fixing, market division and boycotts stay prohibited under the bill. TechRepublic reported the measure "remains in the House Judiciary Committee," where companion legislation from Reps. George Whitesides and Robert Latta sits ahead of the midterm elections. Nicholas Felstead, an assistant director at the Australian Competition and Consumer Commission, wrote in a March paper that "legal uncertainty can act as a powerful deterrent," TechRepublic reported.

Not everyone in Congress is convinced

Rep. Josh Gottheimer, a New Jersey Democrat, told Politico that lab executives do not need new legislation to slow down: "If Dario and Elon are truly worried, they can pump the brakes at their own labs, today," he said, referring to Anthropic chief executive Dario Amodei and xAI founder Elon Musk, according to TechRepublic's reporting.

What S.5105 does not cover

S.5105's five categories are written around external attacks and loss of control, not the slower, harder-to-define scenario Pachocki described: a lab voluntarily pacing itself because its own monitoring tools are losing their grip on a model's reasoning. Nothing introduced in Congress yet grants an antitrust safe harbor for that kind of pact, and the bill's requirement to notify the Justice Department in writing before any coordinated action still leaves the department free to challenge an agreement it decides falls outside the five categories.

OpenAI has not said what specific legal opinion it received, if any, or whether it has opened talks with Anthropic and Google on the shared safety standards Pachocki's essay says the industry needs before any lab will actually slow down. The House Judiciary Committee still holds the bill with the midterm elections still ahead, according to TechRepublic, which leaves the antitrust question OpenAI raised unresolved for as long as the bill sits there.