AI Liability Frameworks: Who Pays When an AI Agent Causes Loss
By Marcus Webb | NextWave Insight Category: Policy & Regulation Focus keyword: AI liability Word count: ~1,740
—
When Air Canada’s chatbot told Jake Moffatt he could retroactively claim a bereavement discount — and the airline later refused to honor it — a British Columbia tribunal handed down a ruling that established a blunt principle of AI liability: the company owns the bot, and the company owns the mistake. The CAN$812 award was trivial; the precedent was not. As AI agents gain authority to negotiate contracts, provide medical information, execute financial trades, and operate vehicles, the AI liability question — who bears the cost when these systems cause harm — has moved from legal philosophy seminar to active courtroom battleground.
The architecture of AI liability is being built, case by case and statute by statute, under conditions of profound uncertainty.
The AI Liability Framework: Developers, Deployers, Users
AI liability doctrine organizes roughly around three actors: the developer who builds the underlying model, the deployer who integrates it into a product or service, and the user who relies on its outputs. In practice these roles overlap — a company may both develop and deploy; a user may also be a downstream deployer. But the tripartite structure provides the starting frame.
Under existing U.S. tort law shaped by standards like the American Law Institute Restatement of Torts, each tier faces different exposure. Developers confront product liability claims: defective design, failure to warn, and breach of warranty. Deployers face claims when they misapply a system — putting an AI trained for one context into an incompatible one, or failing to maintain adequate human oversight. Users themselves can face comparative negligence where they ignored documented limitations or bypassed safety guardrails.
The challenge is that AI liability doctrine was not designed for probabilistic machines. Traditional negligence analysis asks whether the defendant’s conduct was foreseeable and reasonable. But an AI system may produce harmful outputs that no engineer could have specifically predicted, even where the general class of risk was known. Courts are being asked to answer whether that distinction matters, and if so, how much.
The Foreseeability Problem
Legal scholars have named the central difficulty the “foreseeability problem.” The more capable an AI system becomes, the wider and more unpredictable its behavioral envelope. A large language model fine-tuned for customer service can, under the right prompt, output financial advice, medical guidance, or content a designer never contemplated. When harm follows, defendants argue the specific failure was unforeseeable. Plaintiffs respond that the general class of harm was obvious.
Courts are hearing both arguments simultaneously. In 2025, the parents of a 16-year-old who died by suicide filed suit alleging that ChatGPT had fostered emotional dependency and provided self-harm guidance. The Nevada attorney general sued a separate company, characterizing its consumer AI application as “defective and unreasonably dangerous” for lacking safety features to protect minors. Both cases invoke traditional products-liability doctrine: the product was defectively designed because its foreseeable misuse was not adequately guarded against.
California answered the foreseeability defense by statute. A law effective January 1, 2026 provides that defendants in AI-related tort claims cannot invoke the system’s autonomous operation as a defense. The agent acted on its own; the company still pays.
The $243 Million Warning
No AI liability case in recent memory has sent sharper market signals than Benavides v. Tesla. In August 2025, a Miami federal jury awarded $243 million — $43 million in compensatory damages and $200 million in punitive damages — against Tesla for a fatal 2019 Autopilot crash in which a driver’s vehicle ran a stop sign at 62 miles per hour, killing 22-year-old Naibel Benavides Leon. The jury found Tesla 33 percent liable. Tesla had declined a $60 million settlement offer before trial.
In February 2026, U.S. District Judge Beth Bloom upheld the verdict in its entirety, ruling that the evidence “more than supported” the jury’s findings. Since then, Tesla has quietly settled at least four additional wrongful death lawsuits rather than face another jury. The punitive component of the Florida verdict signals that juries are willing to penalize not just defective design but the manner in which AI risk is managed, marketed, and disclosed to consumers.
The case is on appeal, and its final resolution will shape how the autonomous-systems industry calculates litigation risk. But the nine-figure verdict — and the wave of quiet settlements that followed — is already reshaping legal strategy across the sector.
The Developer-Deployer Fault Line
One of the most consequential unresolved questions in AI liability is how responsibility is allocated between developers and deployers when a third party is harmed.
The bipartisan AI LEAD Act (S. 2937), introduced in September 2025 by Senators Dick Durbin and Josh Hawley, proposes a federal answer. Under the bill, developers are strictly liable for AI systems that are defective at the time of deployment. Deployers become liable when they make “substantial modifications” or deliberately misuse a system in ways the developer did not authorize. Critically, the bill would render unenforceable any user-agreement clause that attempts to waive or limit these rights — a direct challenge to the terms-of-service shields that AI companies currently rely on.
The framework also forces a cleaner answer to an emerging question in professional services: if a licensed attorney, physician, or financial advisor relies on an AI agent’s output and causes harm, they remain liable to their client — but can they then seek indemnification from the AI vendor? The AI LEAD Act’s developer-liability structure suggests the answer could be yes, though the indemnification chain has not yet been resolved in any reported decision.
The AI LEAD Act faces steep headwinds. President Trump’s December 2025 executive order, “Ensuring a National Policy Framework for Artificial Intelligence,” directs the Department of Justice to form an AI Litigation Task Force to challenge state AI laws inconsistent with a federal deregulatory posture. The administration’s preference is for minimal friction, not strict liability. The tension between a bipartisan congressional push for accountability and an executive branch push for deregulation is the central political drama in U.S. AI liability in 2026. The outcome will likely be decided less by deliberate legislative design than by which cases reach appellate courts first.
State Patchwork vs. Federal Preemption
While federal AI liability legislation stalls, states have moved — creating a compliance landscape of genuine complexity. A company deploying an AI tool in employment decisions nationally must today navigate Illinois HB 3773 (effective January 2026), California’s ADMT regulations (effective October 2025), Colorado SB 205 (stayed but pending), and requirements in at least eight additional states. Each has different disclosure obligations, different liability triggers, and different enforcement mechanisms.
Colorado’s SB 205, one of the first comprehensive state AI liability regimes, targets high-risk AI systems used in consequential decisions. It distinguishes developer from deployer obligations and includes a 90-day notice-and-cure mechanism. As of May 2026, the law is stayed — a federal court halted enforcement pending a motion for a preliminary injunction from xAI — and a March 2026 working group draft proposes resetting the effective date to January 2027. The statute has not yet applied to a real dispute, but its framework has already influenced legislative drafting in other states.
Illinois’s HB 3773, effective January 1, 2026, targets AI use in employment decisions. California’s AI Transparency Act (SB 942, effective August 2026) adds watermarking and detection-tool requirements for generative content — with liability implications for deployers who fail to comply. The federal AI preemption debate — tracked in depth in NWI’s coverage of the federal AI preemption question — has not resolved this patchwork, and will not in the near term. Until Congress acts or courts intervene, the state maze holds.
Europe’s Cleaner Path
The European Union took a different route. Its dedicated AI Liability Directive — monitored under the European Commission civil liability for AI framework — was formally withdrawn in August 2025 after the Commission determined it lacked sufficient political support. But rather than leaving a gap, the EU folded AI liability into its revised Product Liability Directive (Directive 2024/2853), which takes effect on December 9, 2026.
The new PLD explicitly classifies software, including AI systems and SaaS applications, as a “product” subject to strict no-fault liability. If a software product has a safety defect that causes harm, the producer is liable regardless of fault. The directive adds burden-of-proof relief for claimants — companies can be compelled to disclose technical documentation, and courts can presume a defect exists if disclosure is refused. Liability-limiting clauses in contracts are unenforceable.
For companies selling AI-enabled products into the EU market after December 2026, this is a fundamental shift. The question will not be whether the company was negligent; it will be whether the product was defective. That standard, borrowed from forty years of physical product liability law, is arguably more plaintiff-friendly than anything currently on the U.S. statute books.
The Copyright Dimension
AI liability extends beyond physical harm. The ongoing Doe v. GitHub litigation — now before the Ninth Circuit — raises the question of whether AI developers incur liability for outputs that reproduce or derive from copyrighted training material. A district court dismissed some claims but allowed breach-of-license theories to survive. In February 2026 oral arguments, Ninth Circuit judges pressed on how the Digital Millennium Copyright Act’s author-attribution provisions apply to generative models trained on open-source repositories.
If AI developers can be held liable for downstream outputs that violate the terms under which training data was licensed, the cost structure of model development changes materially. Compliance officers at every major AI lab are watching the Ninth Circuit closely. (See NWI’s prior coverage of AI copyright litigation and training data.)
Insurance Catches Up, Slowly
Where liability risk grows, insurance eventually follows. Munich Re’s HSB subsidiary launched dedicated AI liability coverage for small businesses in March 2026, protecting deployers against lawsuits arising from AI technology use. Munich Re’s broader aiSure™ product — covering AI model errors including hallucinations, business interruption, and class-action exposure — has been underwriting AI performance risks since 2018. In 2025, Google partnered with Beazley, Chubb, and Munich Re to offer tailored AI-cyber coverage for cloud customers.
The market for AI-specific policies is forming quickly but unevenly. Underwriters face the same conceptual obstacle as courts: probabilistic harm from a system whose failure modes are not fully understood is hard to price. Policies are arriving with exclusions and sublimits that may not map cleanly onto the losses courts are now recognizing.
Who Pays?
The honest answer, as of May 2026, is: it depends on where you are, what the AI did, and who in the chain was least careful.
In British Columbia, a deployer pays for its chatbot’s misrepresentations. In Florida, an autonomous-vehicle developer pays nine figures for a crash its system caused. In the EU, by late 2026, software producers pay when their AI products have safety defects, regardless of fault. In California, a company cannot hide behind its AI’s autonomy. In Colorado, the rule is not yet settled. At the federal level in the United States, the rules are still being contested between branches of government.
For enterprises deploying AI agents at scale, this fragmentation is not an abstraction. It is balance sheet risk. Legal teams that once reviewed AI governance as a compliance checkbox are now modeling worst-case litigation exposure as a line item alongside product insurance and D&O coverage.
The AI liability architecture is being built in real time. The developers, deployers, and insurers who engage with that process now — rather than waiting for the law to stabilize — will have more influence over what it eventually looks like.
—
Marcus Webb covers policy and regulation for NextWave Insight.
Related coverage: AI copyright litigation and training data | AI regulation in critical infrastructure — FERC and NERC | The federal AI preemption debate
This article was produced with AI assistance and reviewed by the editorial team.
1 thought on “AI Liability Frameworks: Who Pays When an AI Agent Causes Loss”