Responsabilità dell'IA: chi è responsabile quando un sistema di intelligenza artificiale causa danni?

When an AI system causes harm in the Netherlands, liability always lies with a person or a company, never with the software itself. In practice the claim is directed at the user (the organisation that deploys the system) under tort law, Article 6:162 of the Dutch Civil Code (BW), or at the producer under product liability, Article 6:185 BW, which does not require fault.

Which of the two bears the loss depends on where the failure occurred, what each party knew and what their contract says. This article covers both positions. The criminal-law framework, from functional perpetration to corporate liability under Article 51 of the Criminal Code (Sr), is discussed in our companion article on se un algoritmo può essere parzialmente responsabile.

Why is the AI system itself never liable?

Dutch law knows only natural persons and legal persons, and an AI system is neither. It cannot be summoned, fined or ordered to pay damages, because it owns nothing; proposals for “electronic personhood” have been discussed at EU level but not adopted.

The question is therefore which party in the chain failed in a duty: the developer of the model, the party that turned it into a product, the supplier or integrator, the deploying organisation, or the person who acted on the output. Liability for AI is the exercise of locating the failure in that chain and matching it to a legal basis.

When is the user liable?

The user is liable when it deploys, configures or relies on the system without the care a reasonable operator would take. The organisation that chooses a tool, feeds it data and acts on its output is judged on its own conduct.

What if AI is used to commit an offence?

Using AI deliberately for an offence does not change the offence. A convincing AI-generated invoice used to defraud a company is fraud (oplichting) under Article 326 Sr, a forged document remains forgery under Article 225 Sr, and AI-written code used to break into a system is computer intrusion under Article 138ab Sr.

Manipulated images of a real person can amount to defamation, and sexual deepfakes fall under the sexual offences rules in force since 1 July 2024 (Wet seksuele misdrijven). Persistent AI-assisted harassment can be stalking under Article 285b Sr, which is prosecuted only on the victim’s complaint. Prompts, account logs, outputs and file metadata are the evidence that links the offence to a person; see our overview of Crimini informatici e criminalità cibernetica nei Paesi Bassi.

What does negligent use look like?

More organisations get into trouble through carelessness than through intent. A system bought for one purpose is used for another, its output is treated as a decision instead of a signal, and no one checks the cases in which it is wrong.

Under Article 6:162 BW, breach of an unwritten duty of care is unlawful, judged by what a careful operator would have done. An employer is also liable for its employees’ faults (Article 6:170 BW), and a party that uses a supplier’s model to perform its own contract answers for that auxiliary (Article 6:76 BW).

The EU AI Act (Regulation (EU) 2024/1689) now fills in that duty of care for high-risk systems. Deployers must use the system according to the instructions, assign human oversight to competent people with authority, ensure relevant input data, monitor operation, keep logs and inform workers before using such a system at work. A court assessing negligence will read these obligations as the standard of careful conduct.

Human oversight must be real. A reviewer who approves almost everything the system proposes, measured on throughput, is not taking a decision; this “automation bias” is where oversight most often fails.

Which risks can the user not pass on?

• Towards the injured party, the user is the visible defendant. A customer refused credit or a candidate rejected by a screening tool will sue the deployer; recourse against the supplier is a separate question under the contract.
• Data protection applies independently. The deployer is usually the controller under the General Data Protection Regulation (GDPR), and Article 22 GDPR gives people a right to human intervention in solely automated decisions with legal or similarly significant effects. See our page on the Regolamento generale sulla protezione dei dati.
• Modification changes your role. A deployer that puts its name on a high-risk system, changes its intended purpose or substantially modifies it becomes a provider under the AI Act, with the provider’s obligations.

When is the producer liable?

The producer is liable without fault for damage caused by a defect in its product (Article 6:185 BW). A product is defective if it does not offer the safety a person may expect, taking into account its presentation, its reasonably expected use and the time it was put into circulation; the injured party must prove the damage, the defect and the causal link.

Claims must be brought within three years after the injured party became aware, or should have become aware, of the damage, the defect and the producer’s identity, and the right expires ten years after the product was put into circulation. Product liability cannot be excluded or limited by contract towards the injured party. The producer can rely on a limited set of defences, including the development risk defence.

What changes for software under the new Product Liability Directive?

The revised Product Liability Directive (Directive (EU) 2024/2853) entered into force on 8 December 2024 and must be transposed by 9 December 2026. It applies to products placed on the market after that date; earlier products remain governed by the current Civil Code rules.

• Software, including AI systems, is expressly a product.
• More parties can be liable: anyone who substantially modifies a product outside the manufacturer’s control, the importer, the authorised representative and, in some cases, fulfilment service providers and online platforms.
• Destruction or corruption of data not used for professional purposes becomes recoverable damage.
• A product can become defective after it is placed on the market, for example through a faulty update or a failure to supply security updates.

The directive also eases proof. A court can order the defendant to disclose relevant evidence, and if it does not comply, defectiveness is presumed. Where technical complexity makes proof excessively difficult, defectiveness or causation can be presumed if the claimant shows they are likely.

What can the contract regulate?

Between businesses, the contract is usually the first battleground. The supplier must deliver what was agreed; a promised accuracy, availability or AI Act compliance that is not met is a breach, with performance, damages and termination available after a notice of default.

Limitation clauses are usually effective between businesses, but a court can set one aside if relying on it is unacceptable under reasonableness and fairness (Article 6:248(2) BW), for example in case of intent or deliberate recklessness. See our page on richieste di risarcimento danni for how such claims are built.

Does the AI Act give a right to compensation?

No. The AI Act is regulatory: it sets duties for providers and deployers and gives supervisors enforcement powers, but it does not create a claim for damages. The European Commission has withdrawn its proposal for an AI Liability Directive, so claims run through national tort and contract law and product liability.

The AI Act still matters in court: non-compliance is evidence of negligence, the mandatory technical documentation and logs can be requested in proceedings, and the Act settles who is provider and who is deployer. Our Studio legale specializzato in diritto informatico deals with those classification questions.

What if user and producer both failed?

Most incidents have several causes: unrepresentative training data, an untested integration, use outside the intended purpose and operators who did not know the system’s limits. Under Article 6:102 BW, parties liable for the same damage are each liable for the whole towards the injured party, and settle their shares among themselves afterwards.

The Dutch childcare benefits affair (toeslagenaffaire) illustrates this pattern: an opaque risk classification, decisions without real individual assessment and consequences borne entirely by the people affected. Its aftermath was administrative and political rather than criminal.

What applies to autonomous vehicles, drones and robots?

Physical harm adds strict liability that does not ask whether anyone was careless. Under Article 185 of the Road Traffic Act 1994 (Wegenverkeerswet 1994), the owner or keeper of a motor vehicle is liable for damage to pedestrians and cyclists unless there was force majeure, regardless of whether a driver-assistance system was steering.

Drones fall under the EU drone rules, with operator registration through the RDW and supervision by the Human Environment and Transport Inspectorate (ILT). The possessor of a defective movable thing that creates a particular danger is liable under Article 6:173 BW, also where the fault lies in its software, so the owner of an industrial robot cannot simply point at the vendor.

What should you check as a user of an AI system?

• Keep an inventory of the AI systems you use, who supplied them and which decisions they influence.
• Establish for each system whether you are a deployer or, after modification, a provider under the AI Act.
• Record the intended purpose and stay within the supplier’s instructions for use.
• Give operators real authority to override the system and do not measure them on volume alone.
• Keep logs and incident records long enough to use them in a dispute.

What should you check as a producer or supplier?

• Document risk management, data governance and testing against the population the system will actually serve.
• Write instructions for use that are honest about limitations and error rates.
• Set an update policy stating who supplies security updates and for how long, in view of Directive (EU) 2024/2853.
• Agree liability caps, carve-outs and indemnities that reflect what the system actually decides.
• Arrange a channel for customers to report incidents and record how each was handled.

Cosa possiamo fare per voi in materia di responsabilità legate all'intelligenza artificiale?

Our IT and diritto penale olandese lawyers advise organisations that build AI systems and organisations that use them.

• We determine whether you are a provider or deployer under the AI Act and which duties apply.
• We review and negotiate AI supply contracts, including caps, indemnities and access to logs.
• We assess a claim after an incident and the evidence you need to defend or bring it.
• We assess criminal exposure of the company and its directors when an AI system was misused.
• We draft internal AI policies on human oversight, logging and incident handling.

Sintesi

• An AI system cannot be liable; the user or producer is.
• Users are liable in tort (Article 6:162 BW), with the AI Act setting the standard of care for high-risk systems.
• Producers are liable without fault for defective products (Article 6:185 BW), within three and ten-year limits.
• Directive (EU) 2024/2853 makes software a product and eases proof, for products placed on the market after 9 December 2026.
• Several liable parties are each liable for the whole towards the injured party.

Domande frequenti