AI systems are highly relevant to liability because they can cause damage in ways that do not fit neatly into traditional legal categories, and with their rapid integration into business and daily life, damage caused by AI systems has emerged as one of the major areas of AI litigation. Yet, the EU AI Act does not address civil liability, and with the withdrawal of the EU AI Liability Directive, originally proposed in 2022 in order to harmonize fault-based liability rules across Member States, we are left with a vacuum in EU-wide uniform liability rules for AI.[7] The question remains: how to map out the rules of responsibility when harm is caused by AI?

A key pillar of the liability framework is the revised Product Liability Directive (EU) 2024/2853 (PLD), which explicitly brings software – including AI systems – within the definition of “product.”[8] The PLD is a strict‑liability regime that Member States must transpose by 9 December 2026 and will sit alongside national fault‑based rules. In Hungary, the amendment to the Civil Code has already been adopted on 16 December 2025 with the new rules applying to products put into circulation after 9 December 2026.[9]

The most significant change is that by broadening the scope of “product”, the manufacturer – in a broad sense, including developers, importers and authorized representatives – is liable for harm caused to consumers by defective AI products, without requiring the injured party to prove negligence or breach of duty.[10]

Another novelty is that the revised PLD now includes compensation for the destruction or corruption of data and for non-material harm. The latter, however, is limited to situations where it results from physical damage or personal injury.[11]

Despite some major changes, the basic nature of product liability law, and its consumer-oriented focus will still restrict AI-related claims. PLD only covers damage caused by a defective product[12] and does not apply when the claim is based on other grounds (e.g. negligent human supervision) or where the claim concerns purely financial loss (such as AI‑driven trading losses) or purely psychological harm (for example, reputational harm). In addition, the PLD does not extend to claims where the injured party is not a natural person[13] (e.g. algorithmic errors causing business interruption).

For such claims that fall outside the scope of strict liability, national fault-based liability regimes will apply, putting the victims in a very difficult situation, as national liability rules are not yet adapted to address claims for damages caused by AI. The Hungarian civil liability system is no exception, as it follows the same fault‑based approach thus faces the same shortcoming in handling AI‑related claims: in order to seek damages, the proof of fault and causation is the victim’s burden: they must prove the unlawful act or omission of the defendant in deploying, developing or using the AI and the causal link between such unlawful act or omission and the damage occurred. Victims face serious challenges on both fronts: due to the self-learning capabilities and opacity of AI systems, (i.e the so-called “black box effect”) and their autonomous operation, it is not always clear whether an error or defect can be attributed to the system – and if it can, who is to be held liable: the provider, the deployer or the user – or whether a causal link exists between wrongful behaviour and damage.[14]

Although no examples of tort actions against an AI provider have been identified within the EU to date, this is expected to change once the revised PLD is fully implemented.

Questions of contractual liability typically arise when AI, as the subject matter of a contract, fails to perform as promised or advertised. In practice, this may involve a failure to meet agreed performance benchmarks or delivery timelines, or a defect in the AI’s capabilities.

Under Hungarian law, the claimant bears the burden of proving both the existence of a defect and its presence at the time of performance. Given the well-known characteristics of self-learning AI systems and their continuous evolution, defects may emerge not only during initial creation or programming, but also after deployment or delivery to the end user. However, due to the inherent opacity of such systems, it is often difficult to pinpoint when and where a defect originated, or whether the resulting damage stems from a design or development flaw (such as a malfunction in the underlying algorithm) or misuse or improper implementation by the user.

A separate issue concerns the scope of recoverable damages. Under most contract law regimes, including Hungarian law, recoverable damages are limited to losses that were reasonably foreseeable at the time of contracting. Therefore, losses caused by unforeseeable “black box” errors may not be fully recoverable even where a breach is established, unless the contract expressly allocates that risk.

In light of these considerations, the careful scrutiny of contractual provisions governing AI systems will bear increasing importance. Businesses supplying or providing AI are likely to prevent themselves by excluding or limiting liability for defective AI[15], to an extent made possible by national law.