When AI causes damage: who is liable and who pays
The AI Liability Directive has been withdrawn, new product liability law applies from 9 December 2026. Who is liable, and what insurers are now excluding.
In short
Work can be delegated to AI, responsibility cannot. Software is no vicarious agent under Section 278 BGB (German Civil Code); liability stays with the user and their duties of selection, instruction and supervision. The Commission withdrew the planned AI Liability Directive in 2025. The new Product Liability Directive (EU) 2024/2853 covers software placed on the market after 8 December 2026.
Auf Deutsch lesen: deutsche Fassung
Update of 28 October 2026. The update below says that the Bundesrat, the second chamber of the German parliament, was considering the act in its second reading. On 16 October 2026 it did not call for the mediation committee. The Act to Modernise Product Liability Law has since been promulgated (Federal Law Gazette, BGBl. 2026 I No. 321 of 27 October 2026) and enters into force on 9 December 2026 as planned, Articles 2 and 5 to 16 already on 28 October 2026. The transitional rule stays: whatever is placed on the market up to and including 8 December 2026 falls under the old law according to Section 23. I am leaving the text as it is and putting the reference above it, rather than correcting it silently.
Update of 16 October 2026. When I wrote this, the second and third readings of the German act were still pending. In the meantime the Bundestag adopted the Act to Modernise Product Liability Law on 8 October 2026 in the version of committee recommendation 21/8429 (Bundesrat printed paper 584/26). It is due to enter into force on 9 December 2026, and the Bundesrat is considering it in its second reading. In addition, a correction on the cut-off date, which already applied when I wrote the text and which I overlooked: a corrigendum in the Official Journal of 7 May 2026 (OJ L 2026/90364) rewords Article 2(1) of the Directive. It applies to products placed on the market or put into service after 8 December 2026, and not only after 9 December, as I quote it below. The German act draws the same line: whatever is placed on the market up to and including 8 December 2026 remains under the old law according to its Section 23. That changes nothing about what the text says, but it does move the cut-off date by one day. I am leaving the text as it is and putting the reference above it, rather than correcting it silently.
The question comes up in every seminar, and it almost always comes in the same form. Someone in the second row raises a hand and says: and if the thing messes up, who is on the hook? Usually a few people laugh, because the question sounds so simple. It is not. The answer is still shorter than most people expect.
You can hand work over to an AI. Responsibility you cannot. That is not a moral maxim for the office kitchen, it is a description of the law as it stands, and nothing about it changes on 9 December 2026, when the new product liability law takes effect. What changes is something else: anyone who wants to hold something against you in future has less to prove. And your insurer is currently rereading its terms.
A machine is not a vicarious agent
German civil law has a well-worn route for the question of who answers for someone else’s misconduct. Section 278 of the German Civil Code (BGB) says in essence: anyone who uses another person to perform their obligations is responsible for that person’s fault as if it were their own. The apprentice miscalculates, the business is liable. Everybody knows that much.
With an AI this route leads nowhere, and for a banal reason: the provision presupposes a person. Software cannot act culpably, because in the legal sense it does not act at all. The same applies to Section 831 BGB, the vicarious agent in tort. That, too, is a human being.
Some people draw the wrong conclusion from this, namely that there is a gap. There is not. If attribution via the agent drops out, liability stays where it already was: with you. You owe your customer a performance free of defects. What you use to deliver it is your business. A tool does not relieve you, it is your tool.
The yardstick is then your own care, and in a business that breaks down into three very concrete duties: you must select the tool carefully, you must instruct the people who work with it, and you must supervise its use. That is exactly the structure every business knows from its machines. Nobody would think of saying, after an accident at the milling machine, that the machine was to blame.
Liability does not migrate to the machine. It stays with whoever bought the machine and put it into the process.
The directive many are waiting for is not coming
There was a plan to regulate this specifically for AI. In September 2022 the EU Commission proposed an AI Liability Directive that was meant to complement the AI Act. Its core was an easing of the burden of proof: an injured party who cannot look inside the black box was to be allowed, under certain conditions, to rely on a presumption.
Nothing will come of it. On 11 February 2025 the Commission put the proposal on the list of withdrawn initiatives in its work programme, on the grounds that no agreement was foreseeable. Since then no AI liability law has been in preparation, and none has been announced.
For you that means two things, and only one of them is good. The good part: no further layer of obligations is being added on top. The less good part: general law continues to apply, and general law does not know your black box. Anyone who wants to prove a fault against you must prove it under the normal rules. Anyone who, conversely, wants to defend themselves must show under the normal rules that they exercised care. Both are only possible with records. Without them you are in a poor position.
What really happens on 9 December 2026
The gap that the AI Liability Directive was meant to close is getting smaller elsewhere. Directive (EU) 2024/2853 of 23 October 2024 completely replaces European product liability law from the 1980s. And it does something that was disputed for a long time: Article 4, point 1 expressly counts software as a product, according to recital 13 irrespective of how it is supplied or used.
The cut-off date is constantly misquoted. Article 2(1) says: the Directive applies to products placed on the market or put into service after 9 December 2026. Not to damage from that day on, but to products from that day on. What you buy today remains under the old law, even if the damage only occurs in three years. Anyone introducing systems now should therefore read the date as a dividing line through their own system inventory.
In Germany, transposition is visibly under way. The Federal Government has presented a draft Act to Modernise Product Liability Law, Bundestag printed paper 21/4297 of 25 February 2026. The cabinet adopted it on 17 December 2025, the first reading was on 4 March 2026, the hearing in the Committee on Legal Affairs on 13 April 2026. The second and third readings are still pending at the time of writing. The Product Liability Act of 1989 is not being amended, it is being replaced.
Now the qualification that is missing from most articles on the subject: product liability hits the manufacturer, not every user. If you merely deploy a bought-in system, your supplier is the addressee, not you. It gets interesting at the edge. Anyone who distributes under their own name, anyone who modifies a model so that it becomes something substantially different, in other words anyone who retrains or rebuilds it in depth, can slip into the manufacturer role themselves. Exactly that is currently happening in small and mid-sized companies more often than those involved realise.
The part with the greatest practical effect, however, is neither the software nor the cut-off date. It is disclosure. The Directive gives injured parties a right to have relevant evidence disclosed to them, and attaches presumptions to the detriment of the other side if it refuses. Translated into everyday operations: anyone who cannot show how their system arrived at a decision does not merely lose an argument. They lose the presumption.
And then the letter from the insurer arrives
The fastest change does not come from the legislator at all. It comes from the insurers, and it comes quietly.
In the USA this is now openly visible. The body that issues the standard clauses for commercial general liability there published endorsements in January 2026 with which an insurer can exclude generative AI: no cover for bodily injury and property damage or for certain other losses that are attributable to generative AI. Several insurers have additionally included blanket AI exclusions in policies for directors’ and officers’ liability and employment risks. In the industry this is called the end of silently co-insuring AI.
This cannot be transferred one to one to the German market, and I expressly do not do so. Here the same movement runs more subtly and is more unpleasant for precisely that reason. The word AI rarely appears in the list of exclusions. It appears in the policyholder’s obligations. If your contract requires you to comply with recognised rules of technology and to follow your internal rules, and you have no rule at all for the use of AI, then the path to an accusation of gross negligence is short.
What you can do about it is unspectacular and takes half an hour. Take the current version of your liability and cyber policy terms and search for the words artificial intelligence, automated, algorithmic and machine learning. Compare that with the version you actually signed. The exclusion does not arrive as a letter with a subject line. It arrives as a new version of the terms at the next renewal.
All three routes end at the same question
Contractual liability, product liability, insurance cover. Three areas of law, three different counterparts, and in the end the same question is on the table in all three cases: what can you show? I have experienced this often enough in audits to know the two types of answer by heart.
- Weak: of course we check the outputs. Robust: here is the approval rule, here are the logs for the last three months, and here are the four cases in which someone objected.
- Weak: the provider is certified. Robust: here is our selection decision with the criteria, here are their answers to our questions, and here is where the contract says what happens to our inputs.
- Weak: the staff are trained. Robust: here is the content, here is the list of participants, and the colleague over there will explain the procedure to you without looking it up first.
- Weak: the system runs stably. Robust: here is the indicator we measure monthly, here is the threshold, and here is the case in which we breached it and switched the system off.
An example where this tips over is one I regularly encounter in purchasing. Somebody decides: anything under EUR 5 the system may approve on its own, that saves the sign-off. Sounds reasonable. The next day ten orders for EUR 4.99 go through. Anyone who is out to exploit it takes 4.95, 4.96, 4.97. And the really crafty one takes 4.99 in a different currency, because the check only compares the number and not the amount.
The point here is not the loophole in the rule. The point is that each of these orders is a valid business transaction. The supplier is holding a contract. “It was the AI” is not a ground for rescission that the law recognises. You deployed the system, you drew the limit, and you drew the limit wrongly. Afterwards all three sentences are your responsibility, not the model’s.
Where the standard thought this through long ago
ISO/IEC 42001 is not liability law, and getting certified to it does not give you a free pass. What the standard provides is the order for exactly those records that will be asked for when things get serious. That is less than some promise, and considerably more than it sounds at first.
Four places carry the main load. Annex A.3.2 requires AI roles and responsibilities to be defined and allocated, in other words the answer to the question of who signed. A.6.2.6 requires operation and monitoring of the system, A.6.2.8 the recording of event logs, and A.9.4 requires the system to be used only as intended and documented. That is the place where the limit from the 4.99 example belongs. Added to this is clause 6.1.4 with the impact assessment, the only one that asks outwards: who is actually affected if the system gets it wrong?
What is remarkable about this is the timing. These records are worthless if they are created after the damage. A log that is only written on request is not a log, it is a recollection. That is exactly why I advise businesses to settle the evidence side before the certification question. The certifier comes by appointment. The damage does not.
Back to the hand in the second row. The answer to the question of who is on the hook is: you. Just as before, when there was no AI in the house. The only difference is that in future you will have to prove it, and that you have just under three months left before the new product liability law draws the dividing line. Who in your organisation bears responsibility when the AI is wrong, and in which document is that written down?
Primary sources
Frequently asked questions
Is an AI itself liable for its errors?+
No. Software is not a person, it cannot act culpably and is therefore neither a vicarious agent under Section 278 of the German Civil Code (BGB) nor a vicarious agent in tort under Section 831 BGB. Liability stays with whoever deploys the system and is measured by their own care in selection, instruction and supervision. There is no separate legal personality for AI, and none is planned in the EU.
Is there an EU AI Liability Directive?+
No, and there will not be one. The Commission presented a proposal in 2022, but announced the withdrawal in its work programme on 11 February 2025 and formally completed it on 6 October 2025, because no agreement was foreseeable. General liability law therefore continues to apply, supplemented by the new product liability. The planned easing of the burden of proof for injured parties has lapsed with the proposal.
What changes on 9 December 2026?+
From then, Directive (EU) 2024/2853 applies to products placed on the market or put into service from that day. Software expressly counts as a product in it. The connecting factor is important: what matters is the placing on the market, not the time of the damage. In Germany the Product Liability Act of 1989 is being replaced for this purpose; the new act has been promulgated since 27 October 2026 (Federal Law Gazette, BGBl. 2026 I No. 321).
Does the new product liability also affect us as a pure user?+
As a rule, no. Product liability is directed at the manufacturer, so for bought-in systems at your supplier. However, you can end up in the manufacturer role yourself if you distribute under your own name or modify a system so that it becomes something substantially different, for example through retraining. Irrespective of that, your contractual liability towards your own customers always remains.
Does our commercial liability insurance cover damage caused by AI?+
You have to read that in your current version of the terms, not in the one you once signed. In the USA there have been standard clauses since January 2026 with which insurers can exclude generative AI from commercial general liability. In the German market the shift mostly runs through the policyholder’s obligations: anyone without an internal rule for the use of AI risks an accusation of gross negligence. Check the terms before the next renewal.
Author & expert review: Lars Zimmermann · ISO/IEC 42001 Senior Lead Auditor & Senior Lead Implementer · ISO/IEC 27001 Lead Auditor & Lead Implementer (PECB)
Last updated: 10 November 2026. Researched and reviewed to the best of our knowledge; not a substitute for individual legal advice.
Questions about your own case?
In a free 15-minute intro call we assess where you stand on ISO 42001, ISO 27001 and the EU AI Act, honestly and without a sales pitch.