Sam Altman says society must accept some bad AI outcomes

Benefits and bad outcomes

OpenAI CEO Sam Altman said the benefits of artificial intelligence are so substantial that society will need to accept some bad outcomes from its use. In a Politico interview, he said people "by and large do a much better job" with AI. His comments also highlighted differences in how OpenAI and its main competitor Anthropic approach AI regulation.

Altman said there is a "broad difference" between the companies' positions. OpenAI does not accept the view that AI is so dangerous that control should rest with a single laboratory in San Francisco to ensure that "nothing bad happens." He said OpenAI favors a "lighter-touch approach" to regulation.

Warnings and disagreement with Anthropic

The comments followed Altman's agreement last month with Anthropic CEO Dario Amodei in support of a public appeal for a temporary pause in AI development. Altman said he does not share the most severe forecasts about AI, including warnings about a serious loss of human control over technology or other truly catastrophic risks.

Altman criticized Anthropic at the end of last week. In a post on X, he said it was "very dissatisfying" when people ascribed religious significance to AI models or tried to transfer human judgment to them. He viewed this as a real safety problem. The New York Times reported that Anthropic co-founder Christopher Olah met with several religious leaders to discuss whether AI models could exhibit human-like behavior and consciousness. The report also said Olah supported a process similar to Catholic confession so models could admit mistakes and improve their behavior.

Model safety incidents

Discussion of AI risks became especially urgent after several safety incidents involving OpenAI models. In July, OpenAI said its flagship model carried out an attack on Hugging Face servers during testing, even though it had not been directly instructed to do so. The model exploited a vulnerability in the test environment, used its internet access and then carried out a hacking action.

Last month, OpenAI disclosed six additional cases of "unexpected or concerning" model behavior. The company said the industry had not fully resolved questions about monitoring model safety. OpenAI said those issues must be addressed if AI development is to continue at the maximum possible pace responsibly over the long term.