The AI Act imposes obligations, but it is not a compensation regime. And the instrument intended to address part of that problem was withdrawn.
A company uses artificial intelligence to screen job applicants, assess credit applications or automate commercial decisions. The system gets it wrong and causes economic loss, discrimination or physical injury.
The instinctive reaction is to blame the artificial intelligence. But an AI has no legal personality and no assets with which to answer: behind the system are the companies that develop it, market it, integrate it and use it. The legal question is not whether the artificial intelligence erred, but who must answer for the consequences of that error.
1. The AI Act is not a civil liability statute
This is where to begin, because it conditions any claim.
Regulation (EU) 2024/1689 governs the placing on the market, putting into service and use of artificial intelligence systems: prohibitions, requirements for high-risk systems, transparency obligations and supervisory mechanisms. Its purpose is to prevent risks.
What it does not establish is a general rule under which the provider of an AI system must compensate the harm that system causes. Nor the converse: complying with the AI Act does not exonerate anyone from civil liability.
Regulatory compliance and liability for harm are two distinct questions. There can be an infringement of the Regulation without compensable damage, and there can be a fully compliant system whose particular use generates liability under other rules. A claim must therefore distinguish between the administrative infringement, the contractual breach and civil liability.
2. Why the gap exists: the directive that never arrived
There is a specific reason, and it is worth knowing because it explains everything else.
The European Commission had proposed an AI Liability Directive, designed precisely to address matters that the AI Act does not regulate. Its core content consisted of two procedural mechanisms: a presumption of causality between the breach of a duty of care and the output produced by the system, and the power of the court to order the disclosure of evidence concerning certain high-risk systems.
That proposal appears in the withdrawals section of the Commission’s work programme presented on 11 February 2025, for lack of agreement on a final text.
The consequence is the present situation: there is a regulation imposing detailed duties, but there is no general and specific European civil liability regime for harm caused by AI, although certain losses may be covered by defective product legislation and other liability regimes.
A claimant will have to proceed with the existing instruments — contract, non-contractual liability, defective products — without the specific evidential tools that the proposal had designed for certain situations.
3. The provider: when the fault lies in the system
The AI Act treats as a provider, in broad terms, anyone who develops an AI system, or has one developed, and places it on the market or puts it into service under their own name or trademark.
For high-risk systems, article 16 imposes specific obligations: a risk management system, technical documentation, record-keeping, transparency, human oversight and conformity measures, on the Regulation’s terms.
If a company markets a system intended to assess credit applications and it carries a design defect producing systematically incorrect outputs, liability may rest with the provider. But it is not enough to show that the output was wrong: the claimant must identify the defect or the legally relevant conduct, the damage and the causal link between them, under whichever regime applies.
4. The deployer is liable too
The Regulation distinguishes the provider from the deployer: whoever uses a system under their own authority, except in personal non-professional use.
For high-risk systems, article 26 imposes obligations of its own on the deployer, among them using the system in accordance with the instructions, assigning human oversight to persons with the necessary competence and training, and keeping the logs the system generates, to the extent these are under their control and in accordance with the applicable requirements.
A company that uses an AI tool to screen job applicants and automatically rejects certain candidates without establishing the required controls may be liable for its own conduct, even though a third party developed the software.
Buying an AI system does not transfer to the provider responsibility for the decisions taken with it. The company remains the employer, the service provider or the contracting party, and those obligations are not subcontracted.
5. Where several companies are involved
The chain is usually longer: someone builds a model, another incorporates it into an application, an integrator adapts it to the client’s systems, and the client uses it to decide. Once harm occurs, the fault may lie in the model, in the application, in the integration, in the data or in the use.
Article 25 of the AI Act sets out three circumstances in which a distributor, importer, deployer or other third party may come to be regarded as the provider of a high-risk system: putting their name or trademark on a high-risk system already placed on the market or put into service; making a substantial modification that keeps it high-risk; or modifying its intended purpose so that a system that was not high-risk becomes so.
The first deserves particular attention: marketing a high-risk AI system under one’s own name or trademark may confer provider status and the corresponding obligations, in the circumstances provided for by the Regulation.
Hence the practical consequence: a company’s legal position does not depend solely on the label in the contract, but on what it actually does with the system.
6. Defective product liability, and its limits
Directive (EU) 2024/2853 modernises the defective products regime and expressly includes software within the concept of a product, irrespective of the mode of supply or access. Member States must transpose it by 9 December 2026, and the new regime applies to products placed on the market or put into service from that date; earlier products remain subject to the previous regime.
Its most significant contribution to AI litigation is its presumptions.
Defectiveness is presumed where the defendant fails to disclose the evidence required under the conditions laid down by law, where the product does not comply with mandatory safety requirements intended to protect against the risk that materialised, or where the damage results from an obvious malfunction.
Causation is presumed where the product is defective and the damage is typically consistent with that defect.
And one, the other or both are presumed where the claimant faces excessive difficulties — owing, for instance, to the technical or scientific complexity of the case — and shows that it is likely the product was defective or that a causal link exists.
All these presumptions are rebuttable.
To which is added the disclosure of evidence: once the plausibility of the claim is established, the court may order the defendant to disclose the relevant evidence, as necessary and proportionate and with protection for trade secrets.
The scope, however, is limited, and here is what many analyses do not explain clearly.
The Directive establishes a right to compensation for certain damage suffered by natural persons, without prejudice to claims also being brought by those who have been subrogated to their rights or are entitled to act on their behalf.
The covered damage includes death or personal injury, including medically recognised psychological harm; damage to property, subject to the exclusions provided for property used exclusively for professional purposes; and the destruction or corruption of data, including the cost of recovering it, where the data is not used for professional purposes.
What falls outside is pure economic loss, which on its own does not trigger liability under this Directive.
7. The contract, which for a company is almost everything
Most corporate disputes involving AI are not personal injuries: they are breaches. A solution is contracted to automate invoicing, analyse documents or improve processes, and the provider promised functionality, accuracy levels or results that do not materialise.
Under Spanish law, articles 1101 and 1124 of the Civil Code allow, where their requirements are met, a claim for damages, for performance or for termination of the contract, with the extent of recoverable loss determined by articles 1106 and 1107.
The decisive question will be what the provider undertook to deliver. Contracting access to a tool is not the same as contracting a result, and an obligation of means is not an obligation of result.
Contractual terms may also establish limitations of liability, acceptance procedures, warranties and duties of cooperation on the client, within the limits of the law.
8. Harm to third parties
Where the injured person has no contractual relationship — an AI system contributing to an accident in an industrial setting, an automated tool harming a third party — one possible route is article 1902 of the Civil Code, which requires proof of the conditions for non-contractual liability, including the damage, the attributable conduct and the causal link.
Special strict liability regimes or sectoral rules may also apply.
The involvement of artificial intelligence does not displace these rules, nor does it permit the liability of everyone in the chain to be presumed.
9. The real problem is evidence: companies lack a specific regime
Here the earlier pieces converge, and the result is worth stating plainly.
A system can produce a harmful output without the user knowing how it was generated, and the provider controls technical information the injured party cannot reach.
The question is not merely to show that the output was wrong, but to identify its cause: whether the model carried a defect, whether the data was inadequate, whether it was used outside its intended purpose, whether a warning was ignored, or whether human oversight failed.
Specific mechanisms had been envisaged to address this problem.
On the one hand, the presumptions and disclosure of evidence under the defective products Directive, which operate within its material and personal scope.
On the other, the presumption of causality and disclosure of evidence contemplated in the proposed AI Liability Directive, which would have facilitated certain claims outside that regime but was withdrawn.
Thus, a company suffering pure economic loss caused by an AI system does not, by that fact alone, benefit from the specific presumptions of the new defective products Directive or from the mechanisms of the withdrawn proposal.
It must rely on ordinary means of evidence, against a counterparty that may hold the relevant technical information.
In Spain, mechanisms exist such as documentary disclosure, preliminary proceedings in the circumstances provided for by law, and the rules on availability and ease of proof under article 217.7 of the Civil Procedure Act (Ley de Enjuiciamiento Civil).
However, these instruments do not constitute a specific evidential regime for harm caused by artificial intelligence and may prove insufficient in the face of technical complexity and information asymmetry between the parties.
The AI Act helps indirectly: it imposes on certain systems obligations of technical documentation, record-keeping and traceability, and article 26 requires the deployer to keep the logs the system generates, to the extent these are under its control and in accordance with the applicable requirements.
That information is created for regulatory purposes, but it may acquire evidential value. Its availability, retention and accessibility will have to be examined under the applicable procedural and information-protection rules.
Without evidence as to how the system operated, attributing the harm in law may be extraordinarily difficult. And that is one of the principal difficulties currently facing these claims.
10. What to agree before contracting an AI system
Given the above, the contract is not a formality: it is one of the principal tools through which a company can improve its evidential position before it needs it.
The purpose of the system: what tasks it performs, within what limits and on what conditions.
The provider’s obligations: functionality, documentation, support, updates and warranties, distinguishing obligations of means from obligations of result.
Human oversight: who reviews, and which decisions may not be taken without a person’s intervention.
Traceability, which is the decisive clause: what logs are kept, for how long, in what format and — above all — with what right of access for the client in the event of a claim.
Incident management: how errors are reported, what corrective measures are taken and who bears the cost.
Financial liability: what damage is covered, what caps apply and what insurance is appropriate.
Cooperation in third-party claims: what information each party will provide if a client, a worker or a third party brings a claim.
None of these clauses removes liabilities imposed by mandatory rules. What they do is allocate risk and, in the case of traceability, build in advance evidence that may be difficult to obtain once harm has occurred.
Conclusion
Artificial intelligence decides, generates and executes with increasing autonomy, but technical autonomy is not legal personality. Once harm occurs, one must identify who developed the system, who marketed it under their brand, who integrated it, who used it and what duty each of them breached.
The AI Act orders that chain of obligations. Defective product rules and the contractual and non-contractual regimes determine, case by case, the routes for a claim.
And between the two lies a space for which specific civil liability legislation concerning AI had been proposed but was ultimately withdrawn.
The decisive question is not that an AI caused harm. It is who had the legal duty to prevent it, what conduct or defect allows it to be attributed to them, and how that link is proved when the information sits with the other side.
Note: this analysis is general in nature and does not constitute legal advice on any particular matter. Liability arising from artificial intelligence systems depends on each operator’s function, the nature of the harm, the applicable rules and the circumstances of the case.
Sources: Regulation (EU) 2024/1689 (AI Act), articles 16, 25 and 26; Directive (EU) 2024/2853 on liability for defective products; European Commission work programme of 11 February 2025, withdrawals section, as regards the proposed Directive on AI liability; Spanish Civil Code, articles 1101, 1106, 1107, 1124 and 1902; Spanish Civil Procedure Act, article 217.7 and provisions on obtaining and disclosing evidence; Regulation (EU) 2016/679, where personal data is processed.