In one sentence
There is no such thing as “AI liability” — only the liability of people who have used a tool; and because software is not a person, everything hangs not on the AI’s error but on one’s own breach of the duty of care.
The obvious route leads nowhere
An AI invents a citation, and it ends up in a pleading — who is liable? The reflexive thought runs via the vicarious agent: under § 278 BGB a debtor is answerable for the fault of those he uses in performing his obligation. That fits the trainee, the freelance contributor, the corresponding firm — not software. The provision presupposes a person to whom fault can be attributed — a program has neither will nor negligence in the legal sense.
Liability thereby shifts to where it belongs: to one’s own breach of duty, measured by §§ 280 Abs. 1, 276, 675 Abs. 1 BGB. The reproach is not that the AI was wrong — but that nothing was checked, nothing tested for plausibility, nothing documented. An AI error is no ground of liability; an unchecked one is.
What that means in the client relationship
Professional conduct law has long known this shift: conscientiousness under § 43 BRAO and the duty of objectivity under § 43a Abs. 2 BRAO apply irrespective of the aid used; professional indemnity insurance under § 51 BRAO, too, attaches to the breach of duty, not to the tool. How seriously courts take this is shown by the KG Berlin case: what was criticised was not the AI, but the absence of any check.
As a rule the risks lie in three places:
- Citations that look plausible and do not exist — every citation belongs at the primary source.
- Calculations of periods that a model estimates instead of computing — deterministic routes of calculation are safer.
- Statements of fact that are supposed to come from the file but were added by the model.
Whether your indemnity cover extends without restriction to AI-supported work is a question for your insurer — one to be put before the event.
Why exclusions of liability do not relieve you
Almost every set of terms of use for AI contains far-reaching exclusions of liability. Two points on that. First, such clauses operate only between the firm and the provider — vis-à-vis the client you owe your own performance; an exclusion as against third parties changes nothing about that. Second, they are subject to review of standard terms: §§ 305c, 307 and above all § 309 Nr. 7 BGB set limits, for instance in cases of gross fault. Anyone excluding “all liability” across the board often stands on less solid ground than it sounds; whether a clause holds is a matter for review in the individual case.
What the new EU product liability regime changes
A structural shift is coming: the EU Product Liability Directive (EU) 2024/2853 expressly covers software including AI as a product and opens up a no-fault claim against manufacturers. Two qualifications belong with it: it is a directive, not directly applicable law, with a transposition deadline of 9 December 2026; what the German transposition in the Produkthaftungsgesetz will look like is a moving target — check the position as you read. And even then it is directed against the manufacturer — it does not relieve the firm of its own duty of care.
What follows in practice
The documented human review is not a bureaucratic ritual but the line of defence — not because an authority wants to see it, but because in a dispute the question is whether a check was carried out. Without a trace, that cannot be shown. The precondition is that a tool makes visible what an answer rests on: which sources went into it, what was shortened, which citations are substantiated. That is why we carry references to sources on every output.
This overview replaces neither legal advice nor an assessment of your insurance cover.
→ AI transparency · System limits · Compliance for firms · Pleadings in the lawyer portal
