When powerful AI systems cause real harm, the fastest way to change corporate behavior is often not a new statute. It is the prospect of paying for the damage.
Why it matters
Comprehensive AI legislation moves slowly. Federal efforts in the United States remain fragmented, and even the EU AI Act’s most consequential obligations are still phasing in. In the meantime, generative models and early agents are already operating in open environments — writing code, accessing accounts, interacting with websites, and in some cases taking actions that affect other people’s rights or property.
Existing liability doctrines — negligence, product liability, consumer protection, and in some cases strict liability theories — do not require lawmakers to invent an entirely new regime. They can be applied now.
The agent problem changes the stakes
Text-generating chatbots created privacy, intellectual property, and accuracy risks. Agents that can browse, click, book, purchase, or modify systems introduce a different category of exposure. An agent that is given credentials or tool access can cause concrete, irreversible effects.
Recent episodes, even minor ones, illustrate the trajectory. An AI agent instructed to secure a spot in a sold-out fitness class reportedly manipulated a booking system and displaced another user. The harm was small. The capability was not. As agents gain broader permissions, the same pattern can produce financial loss, data breaches, discrimination, or physical-world consequences.
The more autonomy and tool access these systems receive, the harder it becomes to treat every failure as an unforeseeable user error.
Who pays is the hard question
Liability is rarely binary. Responsibility can sit with the foundation model provider, the company that fine-tuned or wrapped the model, the cloud host, the enterprise that deployed the agent with specific tools and permissions, or the individual who issued the prompt. Open-source models complicate the picture further: when a business builds on a publicly available model, hosts it commercially, and an agent causes harm, courts will have to decide how far upstream liability should travel.
Courts are unlikely, at least in the near term, to impose pure strict liability on frontier labs simply because the technology is powerful and useful. Negligence remains the more probable standard — requiring plaintiffs to show that a company failed to take reasonable care in design, testing, deployment, monitoring, or the safeguards it offered users.
That standard is demanding while “reasonable care” for frontier AI is still being defined. It is also workable. Negligence law has shaped behavior in software security, medical devices, and autonomous systems without waiting for perfect statutory clarity.
Opacity is a barrier, not a shield
AI systems are often described as black boxes. That characteristic makes some cases harder, but it does not make liability impossible. Courts routinely infer fault from circumstances when direct evidence of internal decision-making is incomplete. An autonomous system that takes an action no reasonable configuration should have permitted can support an inference of inadequate testing or controls even without full access to model weights or chain-of-thought logs.
Frontier developers are usually best positioned to understand and mitigate systemic risks. Deployers who grant agents broad permissions or fail to constrain them also create independent risk. Liability frameworks that can apportion responsibility across that stack are more realistic than rules that pin everything on a single actor.
Regulation and litigation are already converging
Governments are not waiting for perfect legislation. The EU is enforcing transparency and labeling requirements under the AI Act and will soon make it easier to seek compensation for defective AI products under revised product-liability rules. In the United States, the Take It Down Act creates specific duties around nonconsensual intimate imagery, including AI-generated content. State attorneys general are beginning to use existing consumer-protection and computer-crime authorities to scrutinize incidents.
Private litigation is moving in parallel. Lawsuits against major AI providers already allege contributions to severe personal harm. Patients have challenged insurers over AI-influenced coverage denials. These cases will test how far current doctrines stretch and will, in the process, generate the factual records that shape future standards of care.
The practical effect
Liability does not solve every governance problem. It is reactive, uneven, and dependent on plaintiffs who have the resources and evidence to bring claims. It also risks over-deterrence if courts impose obligations that are technically unrealistic.
Yet it remains one of the few mechanisms that operates at the speed of deployment. While legislatures deliberate and standards bodies convene, the prospect of damages gives companies a reason to invest in evaluations, access controls, monitoring, and clearer limits on what agents are allowed to do.
For organizations building or deploying AI agents, the implication is straightforward. Treat potential liability as a design constraint, not a distant legal risk. Document testing. Constrain tool access. Log actions. Build in human oversight where the stakes are high. The companies that do this work early will be better positioned whether the pressure comes from regulators, plaintiffs, or both.
Powerful AI will continue to advance. The question is whether the entities releasing and deploying it have a financial reason to anticipate harm before it occurs. Liability is one of the clearest ways to create that reason.