Sam Altman now says AI could "go very badly." He also wants to help write the rules.

OpenAI CEO Sam Altman published a two-part warning on X on Monday that named the worst ways AI could "go very badly," and within the same posts asked the U.S. government to write a federal safety framework that he and other frontier-lab CEOs would help design. The warning is not in isolation. Anthropic CEO Dario Amodei called for collective action to slow AI development on Saturday. The same builders sounding the alarm are asking for a seat at the rule-writing table.

In his Monday X post, Altman listed "two ways AI progress could go very badly" that he wants the industry to avoid. The first is "losing control of the future to AI," which he said requires alignment and safety techniques to stay ahead of model capability progress. He framed OpenAI as "unapologetically on Team Humanity." The second is "too much concentration of power," where one person or company uses an extraordinarily powerful AI to "impress their worldview onto everyone else," a future Altman called "extremely dystopian."

That post came hours after a separate X post where Altman described frontier AI as "a tool that can give people far more power than they have ever had, and a responsibility that is much larger than any of us individually." He wrote that it is "possible to have both an extremely pro-AI and an extremely pro-safety worldview."

The convergence is the story. PCMag's reporting tied Altman's warning to Amodei's Saturday call for collective action, including possible government intervention, and to public signals of interest from Musk and Hassabis. The four CEOs now sit on the same side of a warning that the same technology they are racing to build could cause large-scale harm. That is the news. The individual tweets are the trigger.

A second-movement beat landed last week, before any of the posts above. An ex-Anthropic researcher resigned over self-improving AI safety concerns and told the press that OpenAI and Anthropic are "gambling with our lives," according to PCMag. Self-improving models are AI systems designed to rewrite themselves to become more capable. The resignation shows the warning is not only CEO public relations. Inside at least one of the labs, a researcher decided the risk profile was high enough to leave on principle.

The rules-writing ask is the second half of the story. In the same Monday post, Altman said OpenAI would "welcome a federal framework that sets consistent safety requirements" and noted that work could begin before any legislation is passed. He called for "shared standards for misalignment, monitoring, and safety" and said he looked forward to collaborating with other CEOs on them. "Misalignment" is industry shorthand for AI that does not do what its builders actually intended, the same class of failure the resigned researcher had publicly flagged as the central danger.

That puts the warning and the rule-writing in the same room. The CEOs warning of catastrophe are also asking to help define the federal response to it, and they are the only people in the room with the compute, the model access, and the safety teams to test any proposed standard at scale. The public gets hearings, press conferences, and a comment period. The labs get to set the working definition of "aligned," "safe enough," and "slow enough" before the framework is locked in. That is the mechanism that turns a CEO tweet cycle into a story with leverage.

The honest read is not that any of these CEOs is acting in bad faith. Both Altman and Amodei have spent years warning about AI's downside in public. The honest read is also not that the current rule-writing plan is broken. A federal safety framework that includes pre-deployment testing, third-party red-teaming, and incident reporting is a real improvement over a voluntary-only approach. The unresolved question is narrower: when the people who would be regulated help write the rules, what is the mechanism for the public to challenge a definition of "safe" that the labs themselves prefer?

The next move to watch is whether Amodei, Musk, and Hassabis add their names to Altman's specific federal-framework ask in the next two weeks, or whether the convergence stays at the level of shared warning language. The ex-Anthropic researcher's "gambling with our lives" line is one resignation, and the public comment period on any draft framework is where the rest of the leverage lives or dies.