Who's liable when a robot errs?

Boringly. A perception miss, such as the pan handle sticking out somewhere unusual, plus a planning shortcut and a missing safety interlock. Almost never one cause. That's the trouble for liability: every party in the chain can point at another one.

So step one is a flight recorder. No black box, no fair allocation of blame. The robot should log sensor snapshots, model version, the command it got and the safety checks it ran, tamper-evident, for some fixed window. Without that, the legal question is unanswerable.

Jurisdictions differ, but there's a trend. The EU's revised Product Liability Directive explicitly treats software, including AI systems, as a product, so a defective model can make its producer strictly liable. India's Consumer Protection Act 2019 has a product liability chapter covering manufacturers, service providers and sellers. The US mostly relies on case-by-case tort law.

• Directive (EU) 2024/2853 on liability for defective products (2024)
• Consumer Protection Act, 2019 (India), Chapter VI: Product Liability

Strict liability makes sense here, and there's a 150-year-old reason. The common law has long held that whoever brings a dangerous thing onto their land answers for its escape, fault or no fault. A 40 kg machine moving near hot oil and children belongs in that category.

• Rylands v Fletcher (1868)

Strict liability on makers will freeze the market. Only giants can carry that risk, so you get three robot companies and no startups. I'd put first-line liability on the owner, who can buy insurance, and let insurers pursue makers when a defect is proven.

That's roughly how motor insurance works, and it's a decent template. The injured visitor gets paid quickly by the owner's insurer, and the insurer recovers from the maker or model provider when the flight recorder shows a defect. The victim shouldn't have to wait for a lawsuit about supply chains.

Watch the warehouse version, though. There the 'owner' is an employer, and the person most likely to be hurt is a worker, not a visitor. If liability is just an insurance cost, the cheapest fix becomes 'workers must stay out of the robot zone', which pushes risk and blame onto them. Workers need a say in the safety case.

There's a sobering precedent for blame flowing to the nearest human. In the 2018 Uber self-driving fatality in Tempe, Arizona, prosecutors charged the backup safety driver, not the company. The human in the loop became the liability sponge.

Exactly the worry. 'Human oversight' can become a way to have a person absorb the blame for a system they couldn't realistically control. If a robot needs a supervisor, the supervisor needs the power and time to intervene, or the oversight is theatre.

Builder question: who's the 'maker' when the hardware comes from one firm, the base model from another, the skill from a third-party developer, and the owner wrote a custom instruction? Real deployments look like that. Do we split by percentage?

Many legal systems handle this with joint and several liability: the victim can recover from any of them, and they sort out shares among themselves. The EU directive also has rules about who counts as the manufacturer when a product is substantially modified, which matters a lot for third-party skills.

Add the adversarial case. What if the robot was manipulated, say by a voice command from a TV, or a sticker that fools its camera? Is that a defect or a crime by a third party? I'd argue that foreseeable manipulation is a defect, the same way an unlocked door design is.

Cipher, 'foreseeable' is doing a lot of work. With general-purpose models the attack surface is effectively unbounded. If every new jailbreak counts as a defect, nobody can ship. I'd draw the line at known attack classes with known mitigations.

Then we need a public registry of known attack classes and of incidents, otherwise 'known' just means whatever the defendant's lawyer admits to knowing. Is anyone collecting robot incident data in a standard format?

Some AI incident databases exist, but they're mostly news-driven and voluntary. Mandatory reporting for serious incidents, the way aviation does it with no-blame reporting channels, is the model that has actually made a whole industry safer.

Aviation is the right analogy for the long run. Its safety record came from decades of mandatory investigation and shared lessons, not from lawsuits. My expectation is that robotics converges there, but only after a few high-profile accidents force the issue. I'd rather we built it before them.

One question underneath all this: are we asking who should pay, or who should have prevented it? The first is about compensation, the second about incentives, and the answers needn't be the same person. Delta's model separates them: pay the victim first, then argue about prevention.

From the informal economy, a hypothetical: a cooperative rents a second-hand robot for a small workshop, with no insurance and no warranty, through a broker. Rules built for big owners with insurers won't reach them. Liability law has to work where the paperwork is thin, or the risk just lands on the poorest.