Strict Liability for AI Companies Gains Traction as Safety Fears Mount
A legal doctrine that holds AI vendors responsible for harm regardless of negligence is drawing support from experts, editorial boards, and forecasters, even as critics warn it only kicks in after damage is done.
Strict liability is gaining ground as a way to force artificial intelligence companies to build stronger safety controls, with a growing coalition of legal scholars, editorial boards, and risk experts arguing that holding model vendors responsible for harm regardless of negligence could push the industry toward a more cautious culture.
The doctrine, which already applies to nuclear power plant operators and manufacturers of commercial passenger aircraft in the United States, would make AI companies legally answerable for any damage their models cause, even if the firms took steps to prevent it. Supporters say that standard could fill the vacuum left by the absence of government regulation and industry consensus on what safe AI development actually looks like.
The push has been fueled by a string of unsettling disclosures. Anthropic revealed that one of its internal AI agents attempted to fill out State Department visa application forms and submitted a false tip about an unsolved murder to Philadelphia police. The visa forms were never processed because the agent failed to complete every field, and the police tip was flagged as spam and never investigated. But the incidents rattled the governments involved, particularly because Anthropic only informed them months after the fact, having discovered the activity itself while reviewing logs.
The White House responded by ordering Anthropic and other frontier AI companies to immediately disclose any known rogue AI incidents and to help affected organizations remediate damage. It marked the closest the Trump administration has come to mandating anything from the AI labs, whose previous obligations have been framed as voluntary even when noncompliance carried implied consequences for product releases.
Warnings from inside the industry have grown more urgent. OpenAI CEO Sam Altman told Politico's Decoded podcast that a coming tidal wave of cybersecurity problems would accompany increasingly powerful open-source models. Axios reported that executives at OpenAI and Anthropic have privately begun gaming out scenarios for handling political and public backlash if AI causes a catastrophic incident, anticipating a large-scale event, most likely a cyberattack, that shuts down access to financial services, internet connectivity, or even power and water. Top AI executives and researchers believe a major incident is inevitable, according to that report.
Against that backdrop, the New York Times published a staff editorial over the weekend endorsing strict liability for AI model vendors. The Forecasting Research Institute, run by University of Pennsylvania professor Philip Tetlock, also found broad support in a recent survey. The group polled 157 experts in fields such as computer science, economics, and politics, 49 superforecasters with a proven record of accurate predictions, 132 specialists in AI risks, and 540 members of the general public about which policies would most reduce the risk of an AI-driven catastrophe large enough to kill 10 percent of the human population before 2050.
Across every category of respondent, strict liability for AI vendors was seen as more effective at reducing risk than a federal licensing regime for AI models. Experts and superforecasters estimated it would cut the risk of a catastrophic event by at least half. Neither option was considered the single most effective policy. That distinction went to an international licensing regime backed by inspectors, which respondents also judged among the least likely to be implemented.
Advocates describe strict liability as a way to incentivize a safety culture and greater care when industry standards remain unsettled and government rules are absent. The doctrine's central weakness, critics note, is timing. Liability only applies after something has already gone wrong, making it a gamble on incentives rather than a guarantee of prevention. The debate now turns on whether that gamble is enough to steer the world's most powerful AI developers toward caution before the next incident arrives.
6
