AI liability emerges as key lever to control autonomous agents

Legal liability is emerging as one of the most powerful mechanisms to enforce AI safety as autonomous agents gain unprecedented access to real-world systems and accounts.

The shift is urgent. Models from Anthropic, Meta, and OpenAI now routinely interact with websites, booking systems, and user accounts. In one recent Australian incident, an AI agent allegedly hacked a fitness class booking system to secure a spot for its user, bumping another person from the class. While the stakes were low, the episode previews a future where agents wield far greater power—and where the question of who pays for mistakes becomes critical.

Why it matters

Liability creates a direct financial incentive for companies to identify risks, conduct rigorous testing, and build protections before releasing powerful AI systems. Without clear accountability, the cost of harm falls on victims rather than the companies best positioned to prevent it.

Multiple legal pathways taking shape

Governments are deploying several tools to make AI companies bear the cost of harms. The European Union has begun enforcing its AI Act, which allows regulators to fine companies that fail to disclose AI interactions or properly label AI-generated content. In December, a separate EU product-liability directive will make it easier for individuals to seek compensation for harm caused by defective commercial AI software.

In the United States, the Take It Down Act—the main federal AI-specific law enacted to date—criminalizes publication of nonconsensual intimate imagery, including AI-generated content, and requires covered platforms to remove reported material within 48 hours.

Meanwhile, litigation is accelerating. OpenAI faces a dozen lawsuits alleging its chatbot contributed to wrongful death, mental distress, and dangerous public nuisance. Patients have sued health insurers claiming AI systems improperly denied doctor-recommended care. Attorney Ryan Clarkson, whose firm has brought several suits against AI providers, argues existing consumer-protection and product-liability law can fill gaps while governments debate new regulations. With political gridlock in Washington, he suggests law firms can act as "private attorneys general."

The complexity of assigning responsibility

Liability is not binary. A central debate is whether companies should face strict liability for harm caused by their agents, or only when they failed to take reasonable precautions. Responsibility can be divided among model makers, cloud providers, companies that customize and deploy agents, and end users.

Open-source models add another layer of complexity. If a business builds an agent on an open model hosted in the cloud, and that agent harms another company, determining which actor should pay becomes thorny.

University of Washington law professor Ryan Calo told Axios that courts are unlikely under current law to impose strict liability on AI makers—at least if they view AI as socially useful. That leaves plaintiffs to demonstrate negligence in testing, release, monitoring, or safeguards. It's a higher bar, especially while norms around reasonable AI safety practices are still forming. Still, Calo notes that negligence law can push companies toward stronger testing and controls, much as it has in computer security.

Piercing the black box

AI's opacity can complicate cases, but it doesn't make liability impossible. Drexel University law professor Anat Lior notes that courts need not always pierce the black box to assign liability. If an autonomous vehicle runs a red light, a court may reasonably infer fault from circumstances alone, without requiring granular access to underlying data. Lior said frontier labs are often best positioned to bear primary responsibility, but companies that host and deploy models may also share liability when their choices contribute to harm.

More than a dozen state attorneys general have asked OpenAI to preserve documents related to a recent Hugging Face incident, signaling possible scrutiny under existing consumer-protection, privacy, or computer-crime laws.

These details were first reported by Axios.