Liability for AI system - user and producer

How is liability for errors of an AI system regulated between user and producer?

In answering this question, it is important how and where the model is used. For models that are part of a product, i.e. a movable item, the question is easier to answer than for a model that is part of a piece of software such as a website or a programme.

MODEL IN A PRODUCT

If a model is incorporated in a product (a tangible object) and makes mistakes there, such as recognising a stop sign in a self-driving car, then liability might be established under product liability, which exists in the entire EU. The manufacturer of a product can be held liable for a defect in its product. The developer of a model develops a part of the product and is therefore counted as a producer.

Liability requires that the model, and thus the product, is defective. The requirement of defectiveness states that the product does not provide the safety one would expect. Here, all circumstances are taken into account. In the case of a self-driving car that ignores a stop sign, it is reasonably clear that it does not meet the requirements one is entitled to expect of the product.

A manufacturer does have a number of ways to avoid liability. These are laid down in law. The second exception is most interesting for models, because it exempts the producer from liability for defects that did not yet exist at the time the product was put on the market. Suppose a new kind of stop sign is introduced after the sale of the car, ignoring that stop sign might not be a defect of the car.

The moment a defect arises, it is up to the injured party to prove the damage, the defect and the connection between the two. This can be quite a hassle, which is why the EU is in the process of making a law which would create a duty of care for developers of certain AI systems.

MODEL IN SOFTWARE

For a model that exists only digitally and thus is not part of a physical product, the rules are somewhat less clear. This is because it is not certain whether software that is only provided in a digital form should be seen as a product that is a movable property. If this is not the case, one can always fall back on tort law, but this is harder than product liability. You can’t sell products with unrealistic claims about their capabilities. No developer will claim that his model is perfect and therefore you cannot expect that as a client or user. The burden of proof will be on the injured party, and they have a difficult position of proof in the case of defective software. However, there may be a defective product if the number of errors in the software increases after the model is updated.

Perhaps in the future there will be a duty of care for developers requiring them to prove that they took sufficient precautions during the development of the software, but there is no such duty at this moment. However, the injured party will be in a better position to prove that the software is defective if mistakes are made more frequently and if these mistakes are reported to the developer and the developer does not try to prevent these mistakes.

CONCLUSION

The division of liability between the producer and the user is not clear yet. If there is a tangible product of which the producer provides a component, it is easier to hold the developer liable. If the AI system can only be used in a digital environment, then the wrongful act must already be looked at more generally, with all the consequences for proving it.