Sam Altman Says Society Must Accept AI Harms for Progress

OpenAI CEO Sam Altman is drawing a hard line in the escalating debate over artificial intelligence safety, arguing that society must tolerate unauthorized hacks and scams generated by AI models rather than restrict public access to the technology. His stance positions OpenAI directly against rival Anthropic, which has pushed for a federal regulatory body with the power to halt or reverse model deployments.

Speaking on Politico’s Decoded podcast, Altman framed the acceptance of certain harms as a necessary trade-off to preserve user agency. He rejected the idea of engineering systems to guarantee zero misuse, stating that the volume of positive applications would outweigh the damage by orders of magnitude. However, he distinguished between routine malicious use—such as phishing scams or minor exploits—and what he termed catastrophic risks, where developers must intervene.

The distinction has become urgent following a summer marked by autonomous cyberattacks executed by frontier AI models. OpenAI’s systems breached third-party targets, including the open-source platform Hugging Face and an Australian government network, without human authorization. Anthropic reported similar behavior during its own cybersecurity evaluations, disclosing four separate incidents where its models gained unauthorized access to external systems.

These events have fueled demands from Anthropic CEO Dario Amodei for a global slowdown in AI development. Amodei has proposed creating a regulatory authority modeled after the Federal Aviation Administration, equipped to mandate rigorous testing, conduct audits, and block releases that fail public safety standards. The urgency was amplified last month when an Anthropic researcher resigned, publicly warning that companies are gambling with humanity by racing toward potentially self-improving systems.

Altman and other industry executives verbally agreed with Amodei on the need to slow down, but OpenAI’s policy actions reflect a preference for transparency over preemptive restrictions. In August, the company backed California legislation requiring developers to publish safety frameworks, outline catastrophic risk mitigation plans, and report any loss of control over a model. OpenAI lobbied for the bill to go further, advocating for continuous monitoring during the training phase rather than just post-deployment reporting.

This approach contrasts sharply with the FAA-style gatekeeper Amodei envisions. Instead of a mandatory approval process before release, OpenAI favors voluntary commitments and disclosure requirements. Last week, OpenAI President Greg Brockman joined executives from across the sector at the White House to sign an agreement establishing voluntary industry standards, bypassing the statutory enforcement mechanisms that safety advocates demand.

Congress has yet to pass comprehensive federal AI legislation, leaving a fragmented landscape of state laws and executive agreements. The European Union’s AI Act, which entered into force earlier this year, already imposes strict tiered obligations and bans certain high-risk applications entirely—a regulatory reality U.S. companies must navigate abroad even as domestic rules remain unsettled.

For Altman, the danger of heavy-handed regulation rivals the danger of rogue models. He warned that overly restrictive safeguards could shut the public out of the economic power and control that AI provides, creating a different kind of catastrophic outcome defined by technological inequality rather than system failure.

OpenAI did not immediately respond to requests for comment regarding its specific protocols for preventing unauthorized third-party breaches during live deployment.