AI civil liability: who pays for the damage and which policy responds

With no dedicated EU directive, damage caused by an AI system is handled through national civil law, the new Product Liability Directive, the AI Act and, in Italy, Law 132/2025 and Legislative Decree 160/2026. A reading from the side of those who take on cyber risk: where liability falls, which policies are exposed, where the market is excluding and which questions to bring to renewals.

Analysis · 1 November 2026 · updated 2 November 2026 with Legislative Decree 160/2026 · by cybercoverinsurance

An AI system gets a quote wrong, rejects a job candidate, suggests the wrong treatment or invents a precedent in a legal opinion. The damage is real, but who is liable? For those who take on the risk, the question hides a more practical one: which policy pays, and has anyone priced it.

The short answer: Europe has no single rule on AI liability. Damage is spread across four different sources, and each pushes the risk towards a different policy. The biggest risk for a portfolio is not a new liability but the exposure left implicit in existing covers.

In the text, "Article" means a binding rule; "recital" means one of the introductory paragraphs that explain how the rules should be read.

In brief

The EU AI Liability Directive was withdrawn: the Commission decided on 16 July 2025, and the notice was published in the EU Official Journal on 6 October 2025.
From 9 December 2026 software, including AI systems, is a product under Directive (EU) 2024/2853: the producer is strictly liable towards natural persons.
The AI Act does not regulate compensation, but it sets the obligations whose breach becomes evidence of fault. The rules for Annex III high-risk systems move to 2 December 2027.
In Italy, Law 132/2025 requires professionals to tell clients when they use AI and confirms that liability stays with them.
Since 30 September 2026, Italian Legislative Decree 160/2026 presumes causation when AI Act obligations are breached and lets claimants sue the liability insurer directly.
In the United States the market is excluding the risk: standard generative AI exclusions for general liability have existed since 1 January 2026.

The framework: four sources, no single rule

1. The directive that never came

In 2022 the European Commission proposed an AI Liability Directive (COM(2022)496), with presumptions in favour of claimants and duties to disclose evidence. The Commission decided to withdraw it on 16 July 2025; the notice was published on 6 October 2025 (EAPIL). AI damage is therefore left to the national law of each Member State.

2. National civil law: the Italian example

Without dedicated rules, general rules apply. In Italy, where there is a contract between the injured party and the user of the AI, liability is contractual and the debtor is liable for non-performance unless it proves the cause was beyond its control (Article 1218 of the Civil Code). Towards third parties, Article 2043 applies, which requires intent or negligence. Legal scholars debate applying Article 2050 on dangerous activities, which would put on the user the burden of proving it took every suitable measure to avoid the damage (Studio Cataldi). There is no settled case law yet.

Since 30 September 2026, Italy also has Legislative Decree no. 160 of 9 September 2026, which implements Law 132/2025 (PG Legal). The decree creates no new strict liability, but it shifts the burden of proof. If AI Act obligations are breached, the causal link between the breach and the damage is presumed unless proven otherwise (Article 18). Compliance with the AI Act, even if certified, does not by itself exclude liability (Article 19). Courts can also order the provider or user to disclose the system's logs and technical documentation (Lavorosì).

3. The Product Liability Directive

From 9 December 2026, Directive (EU) 2024/2853 applies to products placed on the market from that date and includes software in the definition of product (Article 4). An AI system falls within this definition. Three points matter most (EUR-Lex):

in assessing whether a product is defective, its ability to keep learning after being placed on the market is taken into account (Article 7);
courts may presume the defect or the causal link when technical complexity makes proof excessively difficult for the claimant (Article 10);
only natural persons can claim; a business that suffers damage stays on contractual ground.

The topic is covered in detail in the analysis of software as a product.

4. The AI Act and Italian law

Regulation (EU) 2024/1689, the AI Act, does not govern compensation. It does set obligations for those who develop AI systems (providers) and for those who use them in their business (deployers). For high-risk systems, deployers must use them according to the instructions, assign oversight to competent people, monitor how they work and keep the logs they generate (Article 26). In a civil case, a breach of these obligations becomes the simplest way to show fault.

Regulation (EU) 2026/1744, which amends the AI Act, moved the dates: obligations for Annex III high-risk systems, such as recruitment and creditworthiness assessment, apply from 2 December 2027; those for Annex I systems, such as medical devices and machinery, from 2 August 2028 (Orrick). The transparency obligations of Article 50, for example telling people they are interacting with an AI system, have applied since 2 August 2026 (Gibson Dunn).

In Italy, Law no. 132 of 23 September 2025, in force since 10 October 2025, adds a rule for the intellectual professions (Normattiva). Article 13 allows AI only as a support tool, with the professional's own intellectual work prevailing, and requires clients to be told clearly which systems are used (Studio Pizzano). The consequence for professional liability is direct: an error produced by an AI system is the professional's error.

Where the risk lands in policies

Professional indemnity and tech E&O. The most exposed cover. A wrong opinion, faulty code or advice partly generated by a model are professional errors like any other. Most wordings do not mention AI: the cover is there until someone disputes it.

Product liability. From 9 December 2026 anyone who builds a model into a product or software sold in Europe is liable without fault towards injured natural persons. For insurers of device, machinery or software makers, the AI system becomes a component to assess like any other.

General liability. Covers bodily injury and property damage to third parties. In the United States, Verisk introduced three optional generative AI exclusions from January 2026 (CG 40 47, CG 40 48 and CG 35 08) for general and products liability (IndependentAgent).

Cyber. Responds when the damage passes through a security incident: data leaking through an AI assistant, unauthorised access through a compromised model, deepfake fraud. It does not respond to the model's error as such, which is not a security incident.

D&O. Comes into play when the use of AI becomes a governance choice: AI Act obligations ignored, untrue statements to the market about the company's own AI capabilities.

The market is moving in two directions

The first is exclusion. According to an analysis by the Center for Strategic and International Studies dated 4 September 2026, more than 60 US property and casualty groups filed AI exclusions in 2026, and a proposed W.R. Berkley endorsement excludes claims involving "any actual or alleged use" of AI (CSIS). The same source reports a Lloyd's Market Association survey finding that insurers know little about how their insureds use AI.

The second is explicit cover. Dedicated products exist: Munich Re has offered cover tied to model performance since 2018, Armilla, with capacity from Lloyd's syndicates, covers errors and hallucinations up to 25 million dollars per insured, and in June 2026 CFC added explicit AI cover to seven product lines (CSIS).

The pattern recalls silent cyber after 2019, when Lloyd's required every syndicate to state whether each policy included or excluded cyber risk. The Italian market has not yet taken this step for AI.

What it means for underwriting

Silence is the main risk. A policy that does not mention AI does not exclude it, but does not price it either. In a PI or tech E&O portfolio, exposure grows with clients' adoption without the premium noticing.

Aggregation comes from suppliers. Many companies use the same few foundation models and platforms. A systematic error or a faulty update of a widely used model can produce similar claims across many insureds in the same period. It is the same pattern as cloud aggregation, applied to a less visible component.

Documentation decides the claim. With the Product Liability Directive presumptions and the AI Act obligations, an insured that keeps no logs, does not know which model version produced a result and has no one assigned to oversight will struggle to defend itself. Governance quality becomes a severity factor, not just a frequency one.

Direct action brings the insurer into court. Under Article 20 of Italian Legislative Decree 160/2026, a claimant can ask the party responsible, even before suing, whether a liability policy covers the damage, with a reply due within 30 days, and can sue the insurer directly up to the policy limit (PG Legal). The insurer can raise policy defences that predate the loss and, if it pays, seek recourse against its insured to the extent the policy would have allowed it to refuse or reduce payment. For a PI or liability policy that does not mention AI, the implicit cover is tested in court with the insurer as defendant.

The professions are the first test. Under Article 13 of Law 132/2025, an Italian professional who did not inform the client starts a dispute at a disadvantage. For insurers of lawyers, accountants, engineers and consultants, the client notice is a document to ask for.

Impact for insurersMedium

AI damage today concentrates on PI and tech E&O with covers left implicit, and becomes easier to prove, in Italy from 30 September 2026 under Legislative Decree 160/2026 and across Europe from 9 December 2026 under the Product Liability Directive; claim volumes in Italy cannot yet be measured.