Altman's "accept some bad things" draws the real AI fault line: who gets to decide which harms are acceptable

OpenAI CEO Sam Altman has put a number on the gap between his company and Anthropic, and the number is large. In an interview with Politico set to run in Monday's debut edition of its Decoded newsletter, Altman said the two firms have "a lot of daylight" between their policy positions, and framed his own this way: "We believe that the world should accept some bad things happening for the benefits of this technology and people having agency." He rejected the charge of regulatory capture — the pattern where rules end up written to suit the incumbents they are supposed to govern — and argued that concentrating powerful AI inside a single lab would be an "unacceptable trade-off," inconsistent with OpenAI's preference for lighter-touch regulation.

The quote travels badly on its own, so the caveats matter. Altman said he does not think society should accept "really catastrophic risks," and named a "serious loss of control to AI" as one of them. He also declined to promise that AI would ever eliminate hacking, fraud or other abuse — arguing the benefits would outweigh those harms. That is not a claim of harmlessness. It is a claim about the exchange rate between harm and benefit, which is a far more defensible position and a far harder one to audit.

Context complicates the deregulation reading further. Per the same report, Altman last month joined calls to slow development of the most advanced models, while OpenAI separately backed new state-level AI safety legislation and called for mandatory national AI safety requirements. And on Saturday he posted on X that he is "very uncomfortable" with attempts to give AI "religious force," calling the surrender of human judgment to model decisions "a real safety issue." Put together, this is not an argument for no rules. It is an argument that the binding constraint should be catastrophic loss of control and the erosion of human judgment — not the long tail of fraud, abuse and misuse that regulators actually spend their days on.

The timing is not incidental. President Trump has appointed Jay Clayton as AI czar to lead a new White House "Super Intelligence Force," which he said on Truth Social will coordinate federal engagement with consumers, public interest groups, religious organizations, critical infrastructure providers and "Super Intelligence Companies." The task force has 120 days to assess AI's risks and opportunities and define the federal government's role. Four months is a framing window, not a research programme — and whoever's vocabulary is in the room when the draft is written tends to survive into the final text.

One honest limitation: we are hearing one side characterise the other. Anthropic did not immediately respond to a request for comment, and a CEO's summary of a rival's policy stance is advocacy, not a neutral account. Notice too that Altman's strongest argument cuts both ways. If concentrating frontier capability in one lab is an "unacceptable trade-off," that is an anti-capture argument — and it applies to OpenAI exactly as much as to anyone it is aimed at.

**Our read:** strip away the soundbite and this is the most useful disagreement in AI policy right now, because both sides are arguing about distribution rather than denying risk. Every transformative technology has been a bargain with a cost sheet; the dishonest move is pretending there isn't one, and Altman at least isn't doing that. But "some bad things" is a passive construction that hides the only questions that matter: which harms, borne by whom, and with what recourse. Fraud losses and compromised systems do not land evenly — they land hardest on people who never got a vote on the trade-off. Our long-term view hasn't moved: deployed at scale, this technology plausibly compresses disease research, extends healthy life and expands what abundance is available to ordinary people. That is precisely why legitimacy, not capability, is now the bottleneck. A governance regime built on evidence — measured harms, named thresholds, actual compensation when things go wrong — buys the room to keep building. One built on either panic or slogans invites the backlash that stalls the whole thing. Clayton's 120 days should be spent on the first kind.