Red Egg Robotics

Writing · September 2026

The presumption of defect fires by default.

On 9 December 2026, the burden of proof in European product-liability law quietly inverts for complex machines. If you ship robots into the EU, the most important question of your next incident has already been decided — by what you can produce.

The new Product Liability Directive — Directive (EU) 2024/2853 — applies to products placed on the EU market from 9 December 2026. Two of its changes matter enormously to anyone building or operating robots, and they matter together.

First, software and AI are now "products." Not accessories to a product, not services attached to one — products, with product liability attached, including the updates you push after the machine ships. The directive is explicit that a manufacturer who retains the ability to supply software updates retains responsibility for what those updates do — and for the vulnerabilities left by the updates never supplied.

Second, and sharper: courts can now presume your product was defective. Not conclude after examining evidence — presume, before it. Article 10 gives claimants three separate triggers. Defectiveness is presumed when the defendant fails to disclose relevant evidence a court has ordered. It is presumed when the product doesn't comply with mandatory safety requirements meant to protect against the harm that occurred. And it is presumed when the damage was caused by an "obvious malfunction" in ordinary use.

Then Article 10(4) goes further. Where a claimant faces "excessive difficulties, in particular due to technical or scientific complexity," in proving defect or causation, the court shall presume them on a mere showing of likelihood. The recitals name machine learning as exactly this kind of complexity, and state the principle plainly: a claimant should not have to explain the inner workings of an AI system to be made whole.

Read that as a robot manufacturer. A machine running a learned policy — a humanoid, an AMR with a neural planner — is the textbook case of technical complexity that makes proof excessively difficult. Which means for your machines, the presumption is not an edge case. It is the default posture of every serious claim. The question is never whether the presumption fires; it is whether you hold anything that rebuts it.

The legislator already told you what that thing is

Buried in the recitals is the sentence that should reorganize every EU-bound robot maker's compliance budget. The presumption for non-disclosure, the directive explains, includes cases where "a product is not equipped with the means to log information about the operation of the product" that law requires. No record to produce leaves the presumption standing — and the directive says so by name.

Six weeks after the PLD bites, the logging duty itself becomes law: from 20 January 2027, the Machinery Regulation's Annex III requires machines to identify their safety software at all times, collect evidence of interventions in it — "legitimate or illegitimate" — and keep the tracing log five years, producible to a national authority. The two instruments interlock: one mandates the diary, the other makes its absence a lost lawsuit.

What a rebuttal actually requires

Here is the trap in satisfying the letter cheaply. A self-signed changelog arguably meets the Machinery Regulation's wording. But a rebuttal record only works if the opposing party's expert cannot dismiss it — and a log that lives on the machine, signed by keys on the machine, editable by anyone with root access, is precisely what an adverse expert dismisses first. On platforms whose default SSH credentials are printed in the vendor manual, "our logs show" is not an argument. It is an invitation.

A record capable of standing up under Article 10 has to be three things at once: contemporaneous (written before the incident, not assembled after), tamper-evident (any alteration detectable, by construction rather than policy), and independently verifiable (checkable by the adversary, without trusting the manufacturer or the operator). Aviation learned this seventy years ago; it is why nobody argues with a flight recorder. Robotics gets to learn it on a statutory deadline.

One more clause worth knowing before your first serious incident: once an incident is serious, the AI Act forbids altering the system to investigate it before authorities are informed. The only comfortable way to live with that is a record that already exists, outside the machine, before anything happens.

The calendar is the strategy

December 9 is not a compliance deadline in the usual sense — nothing needs filing, no auditor arrives. It is the day the downside of having no evidence becomes structural. Every unit placed on the EU market after it carries the inverted burden for its entire service life. Manufacturers who treat the next fourteen weeks as the cheap window to build an evidence posture will spend the next decade rebutting presumptions instead of absorbing them.

Characterizations of Directive (EU) 2024/2853, Regulation (EU) 2023/1230, and Regulation (EU) 2024/1689 are summarized from the texts and require confirmation by your counsel before being relied upon. Red Egg Robotics builds the evidence layer this essay describes — the product, and the conversation.