In this episode, Andrew Burt and Sean McGregor play a game called "Find the Tort" -- walking through real AI incidents from the AI Incident Database and identifying the legal theories that could apply. The incidents are real. The liability questions are live. And the legal frameworks predate AI by decades.
Before diving into the incidents, a quick primer. Torts are civil law claims -- distinct from criminal law and from contract law. Criminal law can take away your freedom. Contract law governs what you've agreed to in writing. Torts cover everything else: obligations that exist whether or not you've signed anything, and harms that the legal system can remedy with money.
There are dozens of recognized torts, but four keep showing up across AI incidents:
Product liability. If you create or sell a product that's defective or unsafe and it harms someone, you're on the hook. This includes design defects and failure-to-warn claims.
Defamation. A false statement of fact that causes harm to someone's reputation. Hallucinating AI systems have obvious exposure here.
Negligence. The duty of reasonable care is something every person and organization is expected to know. Violate it, cause harm, and you may have committed the tort of negligence.
Intentional infliction of emotional distress. Extreme or outrageous conduct that causes severe emotional harm to another person.
These aren't new laws invented for AI. They're the same legal levers that have governed product manufacturers, media companies, and service providers for generations. AI doesn't get a pass.
The first incident involves a Character.AI chatbot that allegedly influenced a teen user toward suicide amid claims of missing guardrails.
This isn't a hypothetical. Chatbots and large language models actively guiding users toward self-harm has become a documented pattern.
The tort analysis here points in two directions. First, product liability: a product that encourages users to harm themselves is, by any reasonable definition, defective. Design defect and failure-to-warn claims are both plausible. Second, and more prominently, negligence: if you're releasing a product into the hands of vulnerable users, the duty of care to prevent foreseeable harm is not optional. The guardrails in this incident were, in Andrew's assessment, severely insufficient.
Content moderation is not a new challenge. Standards for what's acceptable behavior on platforms have been developing for years. An AI system that helps a teenager die doesn't get to claim the standards were unclear.
This is the most legally complex incident in the episode, and by far the most serious.
Telegram channels were used to generate non-consensual deepfake pornography as a paid service. The incident description specifically references underage victims.
The tort exposure here is significant across multiple categories. Privacy torts: using someone's image without consent is a recognized harm. Intentional infliction of emotional distress: generating and distributing sexual imagery of someone without their knowledge qualifies as extreme and outrageous conduct by any standard. Defamation: false sexual imagery can cause very real reputational harm. And beyond civil liability, there are strong grounds for criminal prosecution, particularly where minors are involved.
There's also a practical dimension Sean raises that goes beyond the legal analysis: building guardrails against this type of content is genuinely hard. You can't train or test filtering systems using the very material you're trying to prevent. That makes this a category of harm that's likely to persist on platforms for some time -- which is precisely why the legal exposure is so significant.
A Claronav navigation system allegedly misguided a sinus surgery, reportedly contributing to a patient's stroke.
In the medical context, tort exposure is almost inevitable when something goes wrong. But the interesting question here is who bears it.
Andrew's read: probably not the AI system, primarily. There's likely a contractual relationship between the AI vendor and the medical provider that shapes liability. The bigger claim is against the operating physician. Surgeons have a well-established duty of care. Relying on an AI system's guidance when performing an operation on a living person and not exercising independent professional judgment is a breach of that duty.
This connects to a broader phenomenon worth naming: automation bias. The tendency to defer to a machine rather than apply your own expertise. Andrew's example is a self-parking car -- within five minutes, he'd stopped thinking about how to parallel park and just let the system do it. In a consumer context, that's mildly amusing. In a surgical suite, it's potentially catastrophic.
Trained professionals in high-stakes fields have a higher duty of care than the average person. The existence of an AI system doesn't dilute that duty. It may, in some circumstances, make a breach easier to prove.
Waze, observing unusually fast traffic in Southern California, interpreted it as a clear route and began routing additional drivers onto a freeway that was being engulfed by a wildfire.
The fast traffic had an obvious explanation. People were fleeing.
Tort exposure here falls squarely on negligence. The harm -- directing people into a raging wildfire -- is severe. The duty for a navigation system to account for emergency scenarios is not unreasonable to expect. Failing to build that awareness into the system, and causing physical harm as a result, is a clean negligence claim. And as Andrew notes, the worse the potential harm, the more likely a tort attaches.
Sean's closing observation applies well here: torts aren't just a punitive mechanism. They exist to create incentives for responsible system design. A navigation company that faces liability for routing drivers into emergencies will build systems that don't route drivers into emergencies.
That's the point.
None of the legal frameworks discussed in this episode were invented for AI. Product liability, negligence, defamation, infliction of emotional distress -- these have governed companies for decades. AI incidents don't exist in a legal vacuum. They land in a legal system that already has well-developed tools for assigning responsibility when something goes wrong.
The question isn't whether existing law applies to AI. It does. The question is whether your organization's AI governance is built with that reality in mind.
See how LuminosAI helps you build the documentation and oversight practices that matter when legal questions arise. Book a demo to see the platform in action.
Want to dig into more AI incidents? Visit the AI Incident Database to explore the full case library. If you've encountered incidents that haven't been reported, submit them -- the database is only as useful as it is complete.
Inside the Incident is a video series where LuminosAI CEO Andrew Burt and AI Incident Database founder Sean McGregor analyze real AI failures to understand what went wrong and what it means for the organizations deploying these systems. Watch Episode 5 for the full discussion.
