Liability

From 9 December, physical AI software is a product. Who pays?

3 min readMati Melchior
From 9 December, physical AI software is a product. Who pays?

A physical AI hurts someone. No attacker this time. A sensor misread, the model chose wrong, or an update broke something. Who pays?

In Europe, that answer is about to change. EU member states must transpose the new Product Liability Directive by 9 December 2026. That's 20 days away. After a corrigendum, it covers products placed on the market or put into service after 8 December 2026.

In part 10 of my last series I covered the cyber angle. This post takes the wider one: any defect.

Software is a product. Article 4(1) says "product" includes "electricity, digital manufacturing files, raw materials and software". The code, and the model, sit inside product liability.

Courts weigh a product's ability to keep learning. Article 7(2) lists what counts when judging a defect. Point (c) reads:

the effect on the product of any ability to continue to learn or acquire new features after it is placed on the market or put into service

For a physical AI, that covers fine-tuning and new skills after the sale.

The burden of proof can shift. Article 10 keeps the basic rule. The claimant proves the defect, the damage and the causal link. Then come the presumptions. Article 10(2)(c) presumes a defect where the claimant shows the damage came from "an obvious malfunction of the product during reasonably foreseeable use or under ordinary circumstances".

A court shall also presume defect or causation, or both, in a harder case. Even after the disclosure of evidence under Article 9, the claimant must face "excessive difficulties, in particular due to technical or scientific complexity". The claimant must also show the defect or the link is likely. The maker can rebut every presumption.

A VLA-driven robot looks like a textbook case of technical complexity.

Where do member states stand? CMS tracks transposition. Its November 2025 update named the Netherlands, Sweden and Germany as leading. It also said there was "a lot of work still to do".

The US has no such directive. Its anchor case is a car. In Benavides v. Tesla, a federal jury in August 2025 found Tesla 33% responsible for a fatal 2019 crash of an Autopilot-equipped Model S. The verdict totalled $243 million, including $200 million in punitive damages. In February 2026, Judge Beth Bloom refused to overturn it. Reuters reported Tesla was expected to appeal.

The humanoid matter so far is an allegation. Robert Gruendel, Figure AI's former principal robotic safety engineer, sued in the Northern District of California. He alleges Figure fired him after he warned that its robots could "fracture a human skull". Figure says he was let go for poor performance. It calls the allegations falsehoods.

Here's the honest limit. I haven't read any national transposition law. The CMS update I cite is almost a year old. Benavides is a car case under US law, so it says nothing direct about EU robots. The Figure claims remain unproven. I found no court verdict on a humanoid injury.

Bottom line: for products placed on the market after 8 December, a physical AI's software sits inside EU product liability. Where technical complexity blocks proof, a court can presume the defect. Then the maker has to rebut it.

Physical AI makers: if a court presumed your defect tomorrow, what evidence would you show?

Part 6 of 12 · The Survey: what keeps physical AI safe in 2026.

Share

Physical AI Safety Dispatch

Monthly analysis. No spam. One exclusive insight per issue.

One issue per month. Unsubscribe in one click from any email. Privacy policy.

We use cookies

This site uses essential cookies to function and, with your consent, analytics cookies (Google Analytics) to understand how the site is used. Learn more.