OpenAI asked Congress if slowing down together is legal

OpenAI has asked members of Congress whether frontier labs could legally agree among themselves to slow development down, Decrypt reports on WIRED's story. Chief scientist Jakub Pachocki has called for voluntary slowdowns until developers can show their systems are safe. The question is whether an AI slowdown agreed between rivals breaks antitrust law, and the answer has been sitting in a bill for 51 days.

“The people building AI earnestly believe that it could kill us all by the end of the decade.”

— Jacob Coxon, Former Anthropic engineer, quoted by WIRED

Jacob Coxon, former Anthropic engineer

The law they are asking about

Competitors who agree to limit what they produce are the case American antitrust law treats most harshly:

• Section 1 of the Sherman Act treats an agreement among competitors to restrict output as illegal in itself, with no need to show harm.
• A company convicted under it faces fines up to $100 million, and an individual up to $1 million and ten years.
• The Justice Department spent 154 days investigating four carmakers for a cleaner-than-required emissions deal, then closed the case in February 2020 without acting.

That last case is the closest thing to a precedent. Ford, Honda, BMW and Volkswagen agreed with California in July 2019 to meet tailpipe standards stricter than the federal ones, and the Justice Department opened an antitrust investigation on 6 September. It closed on 7 February 2020 having taken no action. Agreeing to be cleaner than required produced nothing, which is the reassuring part of the precedent, and it still cost five months of federal scrutiny.

The bill that already answers half of it

Senators Adam Schiff and Jim Banks, with Representatives Bob Latta and George Whitesides, introduced a bill on 23 July that carves out exactly this. The Collaboration on Adversarial Threats and Security Risks Act would let AI firms share threat information and jointly delay or limit deployment of high-risk models, with written notice, without antitrust liability. The Justice Department keeps the power to seek an injunction against anyone abusing the exemption, and price fixing and monopolisation stay illegal.

Schiff and Banks wrote the bill around a different threat. It names Chinese distillation of American models and protects coordination against an adversary. Pachocki is asking about something else: a pause among American labs because none of them can prove their own systems are safe as capability climbs toward AGI. The safe harbour covers the first case cleanly and the second one only if a lab can argue its own model is the national security risk.

Why no lab can stop alone

Four people inside the argument describe the same incentive. Anthropic's chief science officer Jared Kaplan has said slowing down alone makes little sense while rivals keep going. Duncan Sabien of the Machine Intelligence Research Institute put the mechanism plainly: every advance yields millions or billions more in funding and hands its creator more power. Miranda Bogen of the Center for Democracy and Technology described companies pushed to ship before they understand the risks.

OpenAI has already done on its own what it is asking permission to do jointly. In August it paused internal work on Astra over cybersecurity concerns, the same model our readers saw building cyberattacks without human help. A pause one company takes alone costs that company. A pause the industry takes together costs none of them, and that symmetry is what antitrust law exists to examine.

Nvidia's Jensen Huang told a Goldman Sachs conference on Thursday that the industry talks about cybersecurity because it is getting ready to sell products, and asked what better way there is to create demand than to create a problem. The labs asking Congress for permission to stop are the same labs whose fear he is describing as marketing. Congress now has to decide which reading the law should assume.

Nothing here should be taken as financial advice; treat it as information to consider.