AI liability: Who pays when AI fails?
AI liability is becoming one of the most contested questions in technology governance. Courts are writing US AI liability law in real time. Every settlement avoided and every verdict appealed leaves a different answer behind.
Not Congress or a federal agency. Courts, through individual cases, have no federal framework to anchor outcomes and no guarantee that the next verdict will go the same way as the last. The EU’s Product Liability Directive requires member states to transpose new AI accountability rules by December 2026. In the US, the answer to who pays when an AI system causes harm depends on which state you’re in, which judge you drew, and whether the company on the other side decided to settle before a precedent could be set.
That’s where things stand heading into what’s shaping up to be a defining year for AI liability.
What recent cases reveal
In August 2025, a Miami jury hit Tesla with a $243 million verdict in a case involving a 2019 Autopilot crash that killed the driver. The case turned on whether Tesla had continued selling Autopilot after becoming aware of a defect, a classic product liability argument that treated the software as the product and the company as the manufacturer. A month earlier, a California judge in a different Autopilot case ruled the software was not at fault. The contrasting outcomes illustrate how differently courts can assess similar AI-related claims.
The cases highlight how unsettled AI liability remains across US jurisdictions.
GM’s Cruise settled for a reported $8 to $12 million after one of its vehicles dragged a pedestrian in San Francisco. Every Waymo case resolved to date has been settled confidentially. No published figure exists. California’s Assembly Bill 1777, effective July 1, 2026, requires autonomous vehicle manufacturers to maintain emergency response lines, equip vehicles with two-way communication devices for first responders, and comply with emergency geofencing directives within two minutes of receiving them, making real-time incident response a legal obligation rather than a voluntary safety measure. Colorado, where Waymo is expanding, has no equivalent. The framework varies by state and continues to evolve through litigation, creating uncertainty for both companies and affected individuals.
The workday case and algorithmic accountability
Derek Mobley applied to over 100 jobs using Workday’s AI screening system, but the system rejected him within minutes each time. In May 2025, he successfully achieved nationwide class action certification for his lawsuit, covering all applicants over 40 who had faced similar rejections by the same system.
The case is less interesting as an employment discrimination case than as a structural one. A human hiring manager who discriminates against older applicants harms one candidate at a time. An algorithm doing the same thing runs across hundreds of employers simultaneously, with no individual person making any decision at any point in the chain.
Cases like this are expanding the debate around AI liability beyond traditional employer accountability.
Existing employment discrimination frameworks failed to address the scale of automated decision-making. Mobley had no contract with the AI developer and had no relationships with most of the employers who rejected him. His only interaction was with a screening system he couldn’t see, whose logic remained undisclosed, and none of the people processing his application personally reviewed it. K&L Gates’ March 2026 analysis describes this situation as plaintiffs reframing “bad outputs” into allegations about AI architecture: they argue not that a decision was wrong but that the system producing it was inherently defective. Continuous bias monitoring aims to catch the structural accountability gap by identifying patterns in production that no individual decision-maker is authorized to create.
Liability challenges in medical AI
When a physician uses an AI diagnostic tool and the output is wrong, who is liable? The question sits at the centre of ongoing discussions about AI liability in healthcare.
Courts have generally held that relying on an AI tool doesn’t transfer the physician’s professional duty of care to the tool or its maker. The doctor who acted on a flawed AI-generated cancer staging result is still the doctor. The 2025 paper in Science examining new case law on medical AI liability puts it plainly: the manufacturer’s exposure depends heavily on whether the failure appears to be a product defect or a foreseeable limitation that clinical judgment should have caught.
The distinction becomes more difficult as AI systems become increasingly capable and autonomous. An AI triage system deployed in a hospital emergency department that systematically underestimates severity for certain demographic groups isn’t failing in a way any one physician chose. The hospital chose to deploy it. The developer built it with that limitation. The physician acted on the output in good faith. When harm occurs to a patient, all three parties are implicated, and none of the existing frameworks cleanly assigns responsibility to all three simultaneously.
The current legal landscape
These issues remain subject to ongoing litigation and regulatory interpretation. The table shows who ends up in court, not who courts have definitively ruled responsible.
| Failure Type | Who Gets Sued | What’s Still Unsettled |
|---|---|---|
| Autonomous vehicle crash | Manufacturer, operator | Federal standards are absent; state laws vary |
| AI hiring discrimination | Model developer, deployer | Class standards for algorithmic harm |
| Medical AI misdiagnosis | Physician, hospital, sometimes developer | Where clinical duty ends and product defect begins |
| AI loan denial | Lender-deployer primarily | Whether the model developer shares liability |
The EU’s Product Liability Directive extends strict liability across AI distribution chains. Developers, deployers, and parties who substantially modify AI systems all face potential exposure under it. California’s Transparency in Frontier AI Act, signed in September 2025, brings civil penalties of up to $1 million per violation for the largest AI model developers. Federal preemption fights over state statutes like this one are expected to reach appellate courts in 2026 and 2027.
Distilled
The evolving landscape of AI liability is being shaped more by courtrooms than by legislation.
Every confidential Waymo settlement is a choice not to let a precedent develop. The Tesla case went to a jury and resulted in a $243 million verdict. Cruise paid somewhere between $8 and $12 million without a trial. Waymo has paid someone, but no one outside the parties knows what or why. That pattern tells you more about the current state of AI liability than any statute does.
The EU’s framework names everyone in the distribution chain and makes them jointly accountable. US law often relies on litigation to determine responsibility among the parties involved. The Workday class action is testing whether algorithmic harm at scale can be treated like a defective product. The medical AI cases are testing whether a physician’s duty absorbs the developer’s responsibility entirely or shares it. Neither question has a clean answer yet.
What changes the picture isn’t more litigation. It’s a federal framework, or an appellate court opinion broad enough to give lower courts something consistent to follow. Both approaches will take time to develop, while affected individuals continue to seek clarity and recourse.
