Invest1 publisher2 min readPublished
OpenAI asks Congress whether rivals could legally agree to slow AI development
WIRED reported the outreach to lawmakers. It comes after a February in which OpenAI and Anthropic both loosened safety commitments, and while a bill from Senators Schiff and Banks would shield some restraint agreements if the Justice Department is told first.
The Investor · Invest desk
What happened
- WIRED reported that OpenAI asked members of Congress in recent weeks whether an agreement among rival AI companies to slow development could violate antitrust law, citing people close to the company.
- Senators Adam Schiff and Jim Banks have introduced legislation protecting certain agreements that restrict AI development to address security risks, provided the Justice Department is notified in advance.
- OpenAI and Anthropic both relaxed safety commitments in February, with Anthropic's Jared Kaplan arguing that slowing development independently made less sense while competitors continued advancing.
- Former Anthropic engineer Jacob Coxon resigned publicly this week and said on X that the people building AI earnestly believe it could kill us all by the end of the decade.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- constraint Without a statute, a lab that wants a slowdown to hold picks between losing position on its own and testing antitrust law alongside its rivals.
- decision Any lab seeking the Schiff-Banks protection would have to tell the Justice Department in advance what it plans to stop building, turning a safety pledge into a filing an enforcer reviews.
- precedent Federal policy so far offers only a voluntary pre-release review of advanced models, so a scoped antitrust exemption would be the first law letting rivals agree with each other about what not to build.
The one estimate on the record of what a slowdown costs comes from a critic of the labs. "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," Duncan Sabien, head of communications at the Machine Intelligence Research Institute, told Decrypt [7]. He also said, "Sans some sort of coordination mechanism, stepping back just means the other guy gets a lead" [8]. Decrypt notes that an agreement among AI companies to slow development could raise antitrust concerns by restricting competition [11].
Miranda Bogen, chief technologist at the Center for Democracy and Technology, 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" [9]. The pressure she describes lands on release dates. A safe harbour would not remove that pressure. It would make a joint decision to miss those dates defensible.
The protection in the Schiff-Banks bill attaches to agreements that restrict AI development to address security risks [3]. Jakub Pachocki, OpenAI's chief scientist, asked for voluntary slowdowns until developers can better demonstrate their systems are safe [2]. Those are two different objects. A demonstration standard covers capability broadly and a security carve-out covers a narrower set of harms, so a lab could win the exemption and still be uncovered for the thing its own chief scientist proposed.
The readings diverge on what OpenAI does with an answer. It may want the shield and intend to use it, in which case February was a response to a legal gap that Congress can close. Or the question is cheap optionality, and a reply that coordination is illegal hands every lab a defensible reason to keep shipping. Or the reply is favourable, nobody files anything, and antitrust was never the constraint that mattered.
I would expect the second or the third. Voluntary restraint needs no clearance from anyone, and of the four slowdown-related OpenAI moves in Decrypt's account, one actually stopped work: the August pause on internal Astra work that lacked stronger safeguards, on cybersecurity grounds [6][16]. Decrypt did not name the lawmakers OpenAI approached [15]. The view fails if OpenAI files an advance notice with the Justice Department, or if it halts a frontier programme before any statute exists to cover it.
What to watch
- Whether any lab files an advance notice with the Justice Department if the Schiff-Banks protection becomes law.
- Whether the bill's scope stays limited to security risks or widens to cover capability agreements as it moves.
- Whether Pachocki's call for voluntary slowdowns turns into a published commitment with dates attached.