Who is liable for a mistake made by AI

The outcome remains human work, even when AI performs the task

An AI system can perform a task, but the system itself is not a legal entity and cannot be held liable. Responsibility for an outcome lies with a natural or legal person: the organisation deploying the system, the supplier of the system, or a combination of both, depending on exactly what went wrong and what agreements have been made about this. This is one of the reasons why ftetoai distinguishes three categories instead of a simple yes or no: for many tasks, the question is not only whether AI can do the work, but who checks which outcome before it goes out or has consequences.

The category "partly, with human oversight" exists precisely for this reason. For these tasks, AI performs the work, but a human approves or rejects the outcome, with a reason. That oversight is not only a quality measure; it is also the point where responsibility for the final outcome is established. Without that oversight, it is often unclear who had which control, and that is exactly the question that becomes relevant once something goes wrong.

Where the rules on liability are laid down

Liability for mistakes made by AI systems is assessed on the basis of existing law: general contract and tort law, product liability, and specific regulation being developed at European level for AI systems with elevated risk. Which rules exactly apply depends on the type of system, the sector, and the way the system has been deployed. This is not an area in which a tool like ftetoai can make a statement that holds true for every situation; the current legal text and its interpretation by a lawyer are the appropriate starting point. Ftetoai describes which part of a task a system can take over and which part requires oversight, not who is legally liable in a specific case.

A separate point is data processing: as soon as a task touches personal data, a second layer of obligations arises. We discuss this in more detail on the page about the GDPR when automating tasks; it describes, among other things, why processing by an AI system requires its own assessment, separate from the question of whether the system performs the task well in substance.

What this means for how the work is organised

Because the outcome remains human work in a legal sense, the way the work is organised is where the practical choices are made. A few recurring mechanisms:

Who checks what, and on what basis. If a task falls into the "partly" category, recording the approval or rejection and the underlying reason is the reference point for reconstructing afterwards what happened.
Which tasks are fully taken over by AI, and under what conditions. For tasks in the first category, there is no human check on a case-by-case basis; the question then becomes how the organisation determined in advance that this is responsible, and whether that aligns with what the supplier and organisation have agreed on this.
How oversight and monitoring are set up at process and team level. This touches on tasks that are themselves partly automatable as well, such as whether AI can take over monitoring budget and flagging deviations or AI can monitor the progress of goals and projects. Here too the same applies: the system signals, a human assesses and decides what happens with it.

This breakdown does not amount to legal advice, but it does provide a factual picture of where the control points in the process lie. That picture is useful in a conversation with a lawyer or advisor about how liability is divided in a specific situation.

Where this touches on internal consultation within the organisation

The way tasks are divided between AI and humans, and how oversight is set up, is not only a legal question but also a question about the organisation of the work itself. Changes to work processes and the deployment of systems that affect employees' work can fall under works council consultation; when that is the case and which steps belong to it is described on the page about when the works council is involved. Here too the same applies: this is a separate track alongside the question of whether AI can take over the task in substance.

A similar mechanism applies to meetings that AI supports rather than replaces. With the question of whether AI can lead the weekly team meeting, it is not only about whether a system can draw up an agenda and take minutes, but also about who remains responsible for decisions made during that meeting. The execution can be supported; responsibility for the outcome does not automatically shift along with it.

What you can do now

This page describes the mechanism, not a ready-made answer for your situation. Liability for a mistake made by AI depends on the type of system, the sector, the agreements made with a supplier, and the way oversight has been set up — and that assessment lies with a lawyer or advisor, not with a tool that categorises tasks.