As an OpenAI whistleblower recently explained, AI executives have knowingly fostered a culture that “guarantees periodic failures,” the scale of which is “growing as systems get more capable.” This approach hinges on an expectation of impunity, which is why AI firms must be held legally accountable for the harm their agents cause.
CHARLOTTESVILLE—Twenty years ago, I argued that the creation of databases of sensitive information—including home addresses, biometric data, and Social Security numbers—should be understood as ultrahazardous activities. They could be managed only if the entities creating them were held fully liable for breaches. As AI models generate new and extreme risks, the need for a strict-liability regime is urgent. If frontier companies bore strict responsibility for the destruction caused by their AI chatbots and agents, they might slow down to figure out how to conduct their activities safely and change course where that is not possible.