ARTICLES

Tort Liability in the Age of Artificial Intelligence

Legal Analysis · 4 min read

Assessing liability for an AI-related loss requires scrutiny of the decisions surrounding the technology: who selected it, what safeguards were required and whether the alleged failure caused the damage.

ByOz Berkovitz

Founder & Owner · Oz Berkovitz Law Office

An inaccurate AI output raises questions about the decisions that led to its use. The more difficult questions concern the people and organisations that designed the system, supplied its information, approved its use and acted on its output. For insurers and claims counsel, the task is to reconstruct those decisions before attributing responsibility.

The starting point under Israeli law

Sections 35 and 36 of Israel’s Civil Wrongs Ordinance [New Version] provide the familiar negligence framework: a duty of care, a departure from the applicable standard of reasonable conduct, and resulting damage. The statutory language expressly addresses the skill and care expected in a profession. Introducing AI into a professional service does not, by itself, answer any of those questions. [1]

In my view, the useful starting point is the function entrusted to the system. Drafting an internal summary and determining whether equipment can safely remain in service present different risks. The assessment should examine the consequences of error, the warning signs available at the time and the checks reasonably practicable before anyone relied on the result.

Identify the decisions that contributed to the loss

Consider a hypothetical claim. A consulting engineer uses an AI tool to summarise maintenance records. The summary omits repeated faults in a fire door. Relying on it without reviewing the underlying entries, the engineer recommends reducing inspections. A subsequent fire causes damage which a functioning door might have limited.

That scenario calls for several distinct enquiries. Did the supplier describe the tool as suitable for safety decisions? Did the implementation team test its treatment of incomplete or contradictory records? What information did the engineer receive, and what checking did the engagement require? Did the building operator implement the recommendation or disregard separate warnings?

The answers may point to failings by more than one participant. Equally, a system’s involvement is not evidence that every participant was negligent. A claim should identify each alleged failure, the duty said to require different conduct and its contribution to the loss. Contractual allocations between suppliers and customers deserve separate analysis from duties potentially owed to the injured party.

Causation requires a record of what happened

Proving that an output was wrong is different from establishing that it caused the claimed damage. In the hypothetical, the enquiry would include whether the omitted information would have changed the recommendation, whether further inspections would have discovered the defect, and whether repair would have reduced the fire loss.

Evidence preservation should extend to the materials underlying the final report. Relevant material may include the source documents, instructions given to the tool, retrieved passages, model version, settings, output history and subsequent human amendments. A later demonstration using a different version may illuminate the mechanism without reproducing the event. Expert instructions should make that distinction explicit.

Responsibility and insurance cover need separate analysis

For an insurer, describing a notification as an “AI claim” does little to identify the applicable cover. The investigation should establish the insured’s actual service, the alleged act or omission, the type of loss and the relevant dates. Those facts can then be tested against the policy’s insuring clause, definitions, exclusions and conditions.

Professional indemnity, product liability and cyber policies may raise different questions. Their potential relevance should be assessed against the particular wording, without assuming that the presence of software makes a claim a cyber event. Supplier indemnities, liability caps and evidence-access provisions should also be reviewed early, including their possible effect on recovery strategy.

A comparative development to watch

The EU’s revised Product Liability Directive expressly brings software within its definition of a product and addresses disclosure of evidence and rebuttable presumptions in defined circumstances. Member States must transpose it by 9 December 2026; its scope, as corrected in May 2026, concerns products placed on the market or put into service from 9 December 2026. As at 24 October 2026, that deadline remains ahead. This is a comparative European development, not a statement of Israeli law. [2] [4]

For businesses supplying European markets, it makes the allocation of technical information and control over updates particularly relevant to contracting and insurance discussions.

Build the evidence before the dispute

A practical review should establish who approves outputs that may materially affect a decision, when concerns must be escalated and which records will remain available after an incident. The Israel Bar’s updated AI guidance illustrates the continuing importance of human responsibility, verification and supervision within legal practice; it is professional ethics guidance, rather than a general rule deciding tort liability in every industry. [3]

Organisations that can explain why a tool was suitable, how its limits were addressed and who made the final decision will be better placed to investigate a loss and present an evidence-based account of their conduct.

Sources

Law and sources reviewed as at 24 October 2026. General commentary only; the applicable law, facts and policy wording must be considered in each matter.