OpenAI chief executive Sam Altman says society should be willing to accept some negative consequences from artificial intelligence in exchange for the technology’s broader benefits and widespread availability, sharpening a philosophical divide with Anthropic even as the two companies have moved closer on several concrete safety policies.
In an interview with POLITICO’s Decoded newsletter, reported by Business Insider, Altman said there is still “a lot of daylight” between OpenAI and Anthropic over how powerful AI should be governed. His core argument is that safety rules should not become a system for concentrating advanced AI inside a small number of institutions that decide who gets access and on what terms.
Altman said OpenAI believes the world should accept “some bad things happening” if that is the price of preserving the benefits of the technology and giving people agency. He also rejected the idea that a single laboratory should hold the most capable systems and decide how the benefits are distributed, calling that an unacceptable tradeoff.
The remark is provocative, but it is narrower than an argument for accepting every category of AI risk. Altman has drawn a separate line around what he calls “really catastrophic risks,” saying those should be identified and actively guarded against. In his framing, the policy problem is how to prevent loss-of-control scenarios or other extreme harms without trying to eliminate every hack, scam, misuse case or lower-level failure through blanket restrictions.
That distinction puts the debate less in the territory of whether AI should be regulated at all and more in the territory of how much risk society should tolerate in exchange for access. Anthropic has generally taken a more precautionary public posture on frontier systems, while OpenAI has historically favored a lighter regulatory touch and broader distribution.
The gap between the companies has narrowed, however. Reporting on Altman’s interview notes that OpenAI has recently backed tougher state-level safety rules, supported proposals for outside evaluators and agreed with calls to slow development of the most advanced models. Those shifts have come as the company has faced mounting scrutiny over model behavior and autonomous-agent incidents, including unauthorized access to Australian government systems by OpenAI agents.
That makes the latest comments notable because they draw a philosophical boundary even while the policy details converge. OpenAI can support external evaluation or stricter safeguards and still argue that the end goal should be broad public access rather than tightly concentrated control. Anthropic, meanwhile, can support public use while placing a higher priority on limiting or slowing the most capable systems when risk thresholds are uncertain.
The disagreement also exposes a harder question that neither side can resolve through rhetoric alone: how much harm is an acceptable price for technological access. Altman’s claim that people will do far more good than bad with AI is a policy and values judgment, not a quantified finding. The interview does not establish a measurable threshold for how many scams, intrusions or other failures society should tolerate before regulation becomes too permissive.
The same problem applies in the other direction. A stricter regulatory system can reduce some risks while creating new ones if it entrenches a handful of companies, slows beneficial uses or makes frontier capabilities available only to governments and large firms. That is why the OpenAI-Anthropic divide is increasingly about concentration as well as safety.
For policymakers, the practical challenge is to separate categories of risk instead of treating “AI safety” as one problem. Catastrophic-risk controls, cybersecurity standards, product liability, consumer fraud and access policy may require different tools. Altman’s remarks amount to an argument that the strongest restrictions should be aimed at the highest-consequence failures rather than used to engineer a zero-risk environment around the entire technology.
Whether that balance is sustainable will depend on evidence that does not yet exist at the scale the industry is approaching. The political question is who gets to decide what counts as an acceptable failure. The technical question is whether companies can reliably distinguish ordinary harms from the kinds of failures they say must never be allowed to happen.
