Sam Altman argues that accepting some harms is part of AI progress

OpenAI CEO Sam Altman argues for less restrictive regulation of artificial intelligence and says society should accept that some harms will occur in exchange for its benefits. In an interview with Politico’s Decoded podcast, he mentioned risks such as hacking, scams and malicious uses, and argued that the technology’s beneficial uses will far outnumber its harmful ones.

His position puts a significant question front and center for companies that develop or adopt AI: what level of risk is acceptable, who decides where the limit lies, and who should be held responsible when something goes wrong. Altman’s remarks express the position of an executive; they do not, in themselves, constitute a rule or determine companies’ legal obligations.

OpenAI and Anthropic disagree on the scope of regulation

Altman described a divide between those, such as OpenAI, who support lighter oversight and those who favor stricter controls. He rejected the idea that control of the technology should be concentrated in a single lab, and argued that imposing measures capable of preventing every incident would mean giving up some of its benefits and limiting its availability.

The comparison with Anthropic is central to the debate, although a company spokesperson told Politico that its regulatory proposals apply to frontier models. The discussion, therefore, is not simply about regulation versus no regulation: it is also about which systems should be subject to specific controls and who should oversee them.

Altman said OpenAI does not accept catastrophic risks, including a severe loss of control over AI. That qualification distinguishes incidents he considers possible and acceptable from the extreme scenarios he says he wants to prevent, although it does not, by itself, establish the criteria for drawing that line.

The debate over risks is also a debate over responsibility

Altman’s remarks drew political and social criticism. Florida Governor Ron DeSantis questioned whether a small group of technology leaders can decide for society as a whole what harms are acceptable. Journalist Chris Stokel-Walker raised a different criticism in an opinion column published by The Guardian: companies should not keep the benefits while shifting the costs of failures onto others.

These are distinct objections, but they point to the same tension. A user who chooses to use a tool may be aware of some of its risks; by contrast, a company, public agency or person affected by a system deployed in their environment could be exposed without having chosen to use it. For organizations, that difference makes it important to assess not only an application’s usefulness, but also who might be affected and what mechanisms are in place to detect, contain and manage incidents.

Altman has also acknowledged the need for a liability framework for companies offering models and said OpenAI will have to take responsibility for at least some problems linked to training. The available sources do not detail how that framework would translate into specific obligations or how responsibility would be divided among developers, providers and user organizations.

A public decision, not just a technological one

The claim that benefits will greatly outweigh harms is Altman’s expectation, not a guarantee established by his remarks. Likewise, acknowledging that incidents will occur does not determine who should bear their consequences. The regulatory debate is precisely about establishing which risks are tolerated, what safeguards are required and what responsibilities belong to those who develop and deploy these systems.

For companies, the debate underscores the importance of treating risk as part of adoption and deployment decisions, especially when a tool can act on third-party systems or data. However, the scope of legal obligations will depend on the applicable rules and the circumstances of each case; executives’ statements do not replace that analysis.