Sam Altman Declares Safety Measures Without Waiting for an Antitrust Exemption

The statement opens on responsibility. "The world deserves confidence that American companies developing increasingly capable AI will act responsibly, especially as the trajectory of progress has steepened. Every frontier lab must deliver on this, and there is no reason any of us should come to work if we cannot," he wrote. He continued: "We welcome a federal framework that sets consistent safety requirements for frontier AI. But we do not believe we need to wait for an anti-trust exemption or legislation to begin the work of providing this confidence." Consistent rules to manage frontier risk are a good idea, he added, and the company is excited by ideas like independent auditors.

There is a backstory to the word antitrust appearing in the statement. METAL reported that OpenAI has spent recent weeks asking Congress whether an industry-wide agreement to slow development would violate antitrust law. David Sacks, the White House AI and crypto czar, responded on X that Anthropic and OpenAI hold a duopoly on frontier intelligence, writing "Stop pretending you need anyone else's permission" and telling them to stop pretending antitrust law has to be suspended so they can form a cartel. METAL also covered Cohere CEO Aidan Gomez's essay rebutting the rules cartel on the same logic. Altman's sentence is an answer to that attack. Not waiting for an exemption means OpenAI can slow its own pace without joining hands with competitors.

The statement starts by admitting the limits of the existing machinery. According to Altman, the Preparedness Frameworks and Responsible Scaling Policies that companies like OpenAI built years ago were good for their moment, but they focused on the deployment of completed models rather than on what happens during development. Now that the focus is shifting to safe development and evaluation, he argued, new tools are needed. His example is the safety case. OpenAI now formulates explicit safety cases in advance of frontier reinforcement learning runs it expects to significantly increase capability, while keeping the safety work it has long done before model releases. Think of an aircraft: instead of filling in the checklist after takeoff, the crew documents why this flight is safe before entering the runway. He said he hopes other companies learn from the approach and propose their own, and that shared standards for misalignment, monitoring, and safety will lead to better outcomes.

He also drew the line himself on what pacing means. "When we talk about 'pacing', we do not mean 'stopping'. Progress has been rapid and will continue to be. But it should be slower than it otherwise could be; interventions like safety cases and monitoring have significant costs," he wrote. He went on: "Pacing will be well worth this cost; no amount of American competitive pressure should justify recklessness, or let capabilities get ahead of alignment and monitoring." Where the company needs government help, he said, is international coordination, but first it should do what it can itself.

Thirteen minutes later he quoted the statement in a second post. There are two ways AI progress could go very badly, it says. The first is losing control of the future to AI, which he called unacceptable: "we are unapologetically on Team Humanity, and AI must always serve people." That requires ways to ensure alignment and safety techniques stay ahead of progress in model capabilities, he wrote. The second is too much concentration of power: if an extraordinarily powerful AI is used by one person or company to impress their worldview onto everyone else, the results could be extremely dystopian. One country gaining too much power and one lab ending up with too much power are the same danger, he wrote, and avoiding both requires walking a narrow middle path. The post had 1.98 million views and 13,000 likes at the same hour.

This statement is the third chapter of a sequence that has run for three days. METAL reported that Dario Amodei, CEO of Anthropic, published an essay on September 12 calling for the frontier to be paced and pledged to give outside evaluators employee-level access. Altman replied on X the same day: "I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks," adding that OpenAI would make the same commitment to independent evaluators with employee-like access. METAL confirmed that this reply stood at 16.11 million views and 66,000 likes as of 9 p.m. on the 14th, the widest reach of any Altman post over the weekend. Elon Musk joined with three words: "Dario is right."

Washington answered differently. According to reports, President Donald Trump told reporters on the 13th, "We're leading China in AI. We're the most sophisticated country in the world, and frankly, I want to keep it that way, because whoever wins AI wins," adding that guardrails can be put in place but that many negative forces are raising things that will not happen. House Speaker Mike Johnson said in a broadcast interview the same day that the three CEOs had privately told him AI needs guardrails but disagreed on what those should be, saying "They don't themselves even have the answer as of today," and that he would not cancel the congressional recess. Altman wrote that he would not wait for legislation less than a day after those remarks.

Seen through a tech lawyer's eyes, the heart of this statement is not a demand for regulation but self-binding. Not waiting for an antitrust exemption means OpenAI will set its pace alone, without agreeing anything with competitors, which is the safest legal position for sidestepping the cartel controversy. In exchange, the promise to document a safety case before training becomes the standard by which the company will later be asked what it knew and what it did if something goes wrong. A safety case is a document the company writes for itself, so the question of who verifies it remains. Altman offered independent auditors and shared industry standards as the answer, but neither yet has a name or a date. China's response went straight for that gap. According to reports, China's Foreign Ministry called the AI risk warnings fear mongering at its regular briefing on the 14th, and state media described Amodei's proposal as a Cold War playbook.

With this statement OpenAI has staked two things on itself. Training runs that sharply increase capability will be preceded by a written safety case, and the results will be shown to independent evaluators and auditors. A company that wrote it will not wait for the law to arrive first now has to prove, on its next training run, that it can keep the promise without one.