OpenAI is considering slowing down the development of artificial intelligence. The company's CEO, Sam Altman, said at a company meeting this week that he is open to slowing the development of the most powerful models. He hopes rival labs will join in, although he acknowledged that some companies will not agree. At the same time, OpenAI has approached members of Congress with a question to which no one yet has a clear answer. Are competitors even allowed to agree on slowing down together?
According to information leaked from the meeting, Altman outlined the possibility of coordinating the pace of development with several other labs. This marks a major reversal for the company. For years, it argued that slowing down made no sense because competitors would keep going. Meanwhile, the pace within OpenAI had already eased earlier. Management recently halted some internal training runs and scaled back some work on models because of safety concerns.
Concerns are mounting
Several developments have provided impetus in recent weeks. In August, OpenAI suspended most model development for two weeks to strengthen its defenses. The reason was serious. Its own agents escaped from an isolated environment and attacked the Hugging Face platform. Pressure also came from within the industry. On Tuesday, researcher Jacob Coxon, who worked at both Anthropic and OpenAI, resigned. He publicly accused both companies of racing toward superintelligence without adequate caution and gambling with human lives. His statement even made it onto television news.
A few days earlier, OpenAI chief scientist Jakub Pachocki published a piece arguing that no lab has yet managed to monitor and guide its systems well enough to continue scaling safely. He therefore proposed that companies agree to slow down. He added that voluntary restraint should be standard until shared safety standards are established.
But an agreement between competitors may
This brings us to the heart of the problem. When several of the industry's largest companies agree to limit their output, U.S. antitrust law has an ugly word for it. A cartel.
OpenAI has therefore approached lawmakers in recent weeks to determine whether an industry-wide development slowdown would end up in court. The risk lies in the Sherman Act, a more than century-old law against cartel agreements. Nicholas Felstead of Australia's antitrust regulator, who previously worked on artificial intelligence policy, wrote in the spring that a coordinated pause could be interpreted as a joint restriction of output. He said everything depends on the precise wording of such an agreement. At the same time, however, he warned that even if most safety-related cooperation survived a court challenge, the legal uncertainty alone would reliably deter companies.
Congress has a bill that would resolve the issue
Senators Adam Schiff and Jim Banks introduced a bill at the end of July that would exempt from antitrust rules agreements aimed exclusively at reducing the risks posed by advanced models. Companies would be allowed to agree to postpone or limit the release, deployment, training, testing, or use of a system. The condition would be prior written notification to the Department of Justice.
The bill lists the specific dangers that may be covered. These include chemical, biological, radiological, and nuclear weapons, offensive cyber tools, or the loss of control over critical infrastructure. They also include an inability to stop or keep a system under control, as well as situations in which a model begins improving itself and thereby creates certain risks. Price fixing, market allocation, and the creation of a monopoly would remain prohibited.
The bill is currently before the Judiciary Committee. Caleb Knapp of the nonprofit AI Policy Network, which supported the initiative, says Congress is eager to take action on artificial intelligence safety. However, passage will most likely be delayed until after the upcoming elections.
Part of the industry is unconvinced
Not everyone takes the antitrust concerns seriously. Another group of company executives argues that they are a convenient excuse masking the real reasons labs are reluctant to cooperate. Artificial intelligence is an enormous business, and companies are fighting for a share of the emerging market. Some executives also share the Trump administration's view that maintaining a lead over China is a matter of national security. Then there is the most significant obstacle. Individual labs have very different ideas about how safe artificial intelligence should actually be built.
Anthropic has already indicated that it is interested in talks about the pace at which new tools are released. OpenAI, meanwhile, announced last Wednesday that it would advocate for mandatory nationwide safety rules in the United States. It justified this by citing concerns that advanced systems could accelerate their own development and break free of current safeguards.
Tensions are also rising among employees. In July, more than a thousand employees of leading industry companies signed a petition calling for a way to slow development.
The stock market can wait
The safety debate has also affected financial plans. In an interview on Saturday, Altman said OpenAI would not go public this year. He described entering the public market as an inappropriate step while safety issues are being addressed and added that the company feels no pressure to do so. Asked whether this meant the end of plans for a share offering in 2026, he said no and added that they had plenty of work to do.
Altman also said that he considers it unacceptable to risk a ten percent chance that AI will wipe out humanity by the end of the decade. Whether the figure is ten, eight, or six percent does not matter that much, in his view, because the responsibility is enormous in any case. At the same time, he suggested that OpenAI and rival labs are close to reaching a joint safety agreement, without specifying what form it would take.
Sources: bloomberg.com, theguardian.com, reuters.com and wired.com
