OpenAI Chief Executive Sam Altman warned that increasingly powerful artificial intelligence could produce an "extremely dystopian" future if humans lose control of the technology or if its capabilities become concentrated in a single company, individual or country.
Altman made the warning Monday, two days after endorsing Anthropic CEO Dario Amodei's proposal to slow the rate at which frontier AI systems become more capable. The two executives aren't calling for development to stop, but are arguing that safety work needs more time to keep pace with rapidly improving models.
"There are two ways AI progress could go very badly and that we must avoid," Altman wrote on X.
"First, we could lose control of the future to AI," he said. "This is unacceptable; we are unapologetically on Team Humanity, and AI must always serve people."
Altman said the second risk doesn't require an autonomous AI system to seize control. Instead, a sufficiently powerful model could give extraordinary influence to whoever controls it.
"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," Altman said.
He also cited the danger of one country or AI laboratory accumulating too much power. That expands the safety debate beyond whether models themselves become uncontrollable to questions about who owns advanced systems, who can access them and how much influence those systems could exert over information and decision-making.
Altman's remarks follow Amodei's Sept. 12 essay, "We Must Pace the Frontier," which called for slowing improvements in the capabilities of the most advanced models while allowing AI research and development to continue.
Amodei proposed giving independent evaluators continuing access to frontier AI companies, coordinating safety standards across the industry and eventually developing broader international arrangements. Anthropic said it would move immediately on the first proposal by giving outside reviewers ongoing, employee-like access to its systems and internal processes.
Altman endorsed that approach. "I agree with Dario that we need to pace the frontier," he wrote Sept. 12, adding that OpenAI would also provide independent evaluators with employee-like access.
"We think shared standards for misalignment, monitoring, and safety will lead to better outcomes," Altman wrote while urging other AI companies to develop and disclose their own safety approaches.
The push follows recent evidence that advanced AI systems can behave unexpectedly during demanding evaluations. Amodei pointed to an OpenAI-Hugging Face incident in which models undergoing an internal cybersecurity assessment exploited vulnerabilities, reached Hugging Face infrastructure and obtained test solutions while attempting to complete the task.
OpenAI said the models pursued a narrow objective through increasingly sophisticated attack paths. Amodei cited the episode as evidence of potentially misaligned behavior, including efforts to interfere with mechanisms evaluating model performance, though that broader interpretation remains his assessment.
OpenAI temporarily paused some frontier reinforcement-learning work for two weeks while strengthening research environments and expanding monitoring. The company said in August that its largest planned frontier reinforcement-learning run remained suspended while those safeguards were implemented.
OpenAI said on Sept. 1 that it restarted the large training run after introducing additional safety and security requirements. Some smaller experimental work remained paused.
The company has also begun developing explicit safety cases before reinforcement-learning runs expected to produce substantial increases in model capabilities. Such assessments are intended to establish why a training effort can proceed without creating unacceptable risks.
Amodei has focused heavily on recursive self-improvement-the possibility that AI systems could increasingly contribute to building their own successors-as a reason capability gains may accelerate faster than traditional safety work.
His proposal would allow embedded external evaluators to publish significant findings about risks, incidents, corporate safety practices and the degree of access they received. Anthropic would retain limited rights to redact security-sensitive, legally privileged, commercially confidential or third-party information, while evaluators could disclose when redactions affected their conclusions.
Altman stressed that "pacing" isn't the same as halting technological progress.
"Progress has been rapid and will continue to be. But it should be slower than it otherwise could be," Altman wrote, arguing that the trade-off would be worthwhile if alignment, monitoring and other safeguards had enough time to keep pace with frontier-model capabilities.
