Sam Altman’s AI risk argument, and the limit he put on it

The trade-off he described

Asked about differences with rival Anthropic, Altman said society should “accept some bad things happening” in exchange for AI’s benefits and people’s agency. Reuters reported that he opposed concentrating access in one San Francisco laboratory, even if the aim was to eliminate misuse, scams and hacks. He expected beneficial uses to outweigh harmful ones. [1]

That is a position about distributing a powerful technology and managing its risks. It is not evidence that any particular model is safe, or a measurement of how often harm will occur. [1]

A boundary around catastrophic risk

In the official transcript’s 8:45–9:15 passage, Altman separates risks that are limited and understood from a serious loss of control to AI. He rejects accepting the latter. Omitting that boundary changes the scope of his argument, making a qualified position sound like acceptance of every possible consequence. [2]

When the interviewer then asks whether Anthropic wants exclusive control, Altman does not affirm the allegation. He says the company would need to answer for its own intentions. His hypothetical criticism should therefore not be presented as a demonstrated account of Anthropic’s plans. [2]

He subsequently describes his preference for concentrating restrictions and testing on the most advanced models, rather than imposing a licensing regime on earlier generations. [2]

What Anthropic’s own proposal says

Anthropic’s July 2025 transparency proposal provides a direct account of one part of its approach. It calls for obligations on the largest developers of the most capable systems, with exemptions intended to avoid burdening smaller developers. [3]

The proposed requirements include publishing a framework for assessing and reducing serious risks, describing tests and mitigations in system cards, and preventing false compliance statements. Anthropic says the framework should remain flexible as evaluation methods evolve. These are the company’s published proposals, not proof that all developers comply with them or that every policy disagreement has disappeared. [3]

The dates behind the discussion

POLITICO promoted the written interview on October 4. Its official podcast listing dates the accompanying Decoded episode to October 5. Those are publication dates; they do not establish the interview’s recording date. This article revisits that discussion rather than presenting the remarks as a new October 13 announcement. [4] [5]