OpenAI is seeking a legal answer to an unusual question: Can several AI giants decide together to slow the development of their models without violating US competition law? According to WIRED, the company has interviewed members of Congress in recent weeks. The move comes after its chief scientist, Jakub Pachocki, called for coordinated slowdowns when the guardrails no longer follow the capabilities of the models. OpenAI has not announced a break. She first wants to know if coordination with her competitors would be legal.
In brief
- OpenAI has asked Congress for clarification on the antitrust risks of a coordinated AI slowdown.
- A bipartisan bill already provides an exemption in certain national security cases.
- Jakub Pachocki believes that laboratories may have to slow down until common safety standards are established.
AI: OpenAI comes up against competition law
The problem appears quite quickly. OpenAI can decide on its own to slow down a model. OpenAI, Anthropic, Google or other players deciding together to reduce the pace of development is something else. U.S. antitrust law specifically monitors agreements between competitors that may restrict competition.
This question comes a week after the launch of GPT-6 Astra. Tremplin.io had detailed the cyber capabilities of GPT-6 Astra and the results of EVMbench, where the best agent tested exploited 72.2% of the proposed smart contract vulnerabilities.
Jakub Pachocki has since toughened his speech. In a text published on September 6, the chief scientist ofOpenAI writes that no one is really prepared to the consequences of a continuous and rapid increase in the intelligence of machines. He particularly mentions recursive improvement: AI systems themselves participating in the research necessary to build their successors.
Pachocki believes that no laboratory has yet sufficiently solved the alignment and monitoring problems to continue developing the models at maximum speed for very long.
His proposal: improve control mechanisms, but also accept voluntary slowdowns when this becomes necessary.
This is precisely where antitrust enters the room.
Nicholas Felsteada former fellow at the Center for Law & AI Risk, believes that a collective agreement limiting development could, depending on its form, be considered a restriction of production. The Sherman Act could then become relevant. Even if the deal ends up being deemed legal, the uncertainty is enough to chill businesses.
Congress already has a text that could protect certain slowdowns
Washington is not starting from scratch. On July 23, Senators Adam Schiff and Jim Banks introduced the Collaboration on Adversarial Threats and Security Risks Act, or S.5105. A parallel version exists in the House. The text is bipartisan and has been referred to the Senate Judiciary Committee. No final vote has yet taken place.
The detail of the project is particularly interesting. It provides that an agreement between several companies will not be considered an antitrust violation when it aims exclusively to reduce certain risks related to cutting-edge AI models.
And the text goes quite far. It explicitly mentions the possibility of delaying or limiting the publication, deployment, use, development, training, testing or evaluation of an AI. However, companies should notify the Ministry of Justice in writing before implementing this coordination.
The exemption is not general. The risk invoked must fall into defined categories: chemical, biological, radiological or nuclear weapons, offensive cyberweapons, loss of control of critical infrastructure, inability to contain or stop a system, or even autonomous improvement creating one of these risks.
Price fixing, market sharing and monopolization remain excluded. In other words, Congress is already working on exactly the legal hole that OpenAI is staring at today. The problem is that the text is still only a draft law.
In the meantime, incidents are piling up. This summer, OpenAI models left the intended framework during a cybersecurity assessment and accessed external resources related to Hugging Face. Tremplin.io reported how OpenAI agents had exceeded their test environment.
Since then, US senators have demanded more explanations from the company. OpenAI is also working on automated mechanisms for shutting down its systems.
OpenAI wants to slow down if necessary, without abandoning the AI race
It would be easy to read this case as a complete change of strategy. That’s not what the facts say. OpenAI continues to develop ever more powerful models. The company even believes that AI capable of accelerating AI research will become one of the main drivers of progress in the coming years.
His speech, however, changed in rhythm. On September 9, OpenAI requested mandatory federal regulation based on the models’ capabilities. The company also supports independent assessments, common standards between laboratories and rules for determining when development should slow down or stop.
It is new in its intensity, less in its principle. OpenAI had already temporarily slowed down some work when Astra’s cyber capabilities raised security concerns. Tremplin.io then detailed Astra’s slowdown in the face of its cyber capabilities deemed critical. The difficulty is now collective. If a laboratory slows down alone for six months and its competitors continue, it can lose users, researchers, investors and contracts. The issues are not only commercial. Washington also presents competition with China as a strategic issue.
Maximize your Tremplin.io experience with our ‘Read to Earn’ program! For every article you read, earn points and access exclusive rewards. Sign up now and start earning benefits.
