Evidence in AI Disputes - November 2026

As AI becomes increasingly embedded in business processes and decision-making, disputes involving AI systems are likely to raise new challenges around evidence. AI outputs may be difficult to reproduce, relevant information may sit across several organisations, and understanding how a system reached an outcome can require highly technical and commercially sensitive evidence.

In the latest session of our AI x Dispute Resolution Webinar Series 2026, William Dunning, Eleanore Di Claudio, and Georgie Courtois considered how these challenges may play out in the English High Court, the UK Competition Appeal Tribunal (CAT), and French litigation.

Why is evidence different in AI disputes?

AI systems can make it more difficult to identify, preserve and interpret the evidence needed to understand what happened. The same prompt may not always produce the same output, while models, configurations and datasets can change over time. Relevant logs may also be routinely overwritten, potentially making it difficult to recreate the system as it operated at the time of the disputed event.

There is also the "black box" problem. Even where an organisation can identify the output produced by a system, understanding why it produced that output may be more difficult. This can create challenges when trying to establish issues such as causation and responsibility.

The evidence may also be fragmented. Developers, providers, deployers and users can each hold different information about how a system was developed and operated. Establishing the full picture may therefore depend on evidence held across several organisations.

Documentary evidence: what you need to know

AI disputes could significantly expand both the types of evidence that need to be considered and the way that evidence is collected and reviewed.

• Preserve evidence early: relevant material could include model versions, prompts, outputs, logs, datasets, source code, model weights and system configurations. Organisations should consider preservation before routine updates, retention policies or deletion processes alter the available evidence.
• Expect different types of evidence: conventional keyword searches may not be sufficient where relevant evidence consists of structured datasets, technical logs or model artefacts. Identifying relevant material may require specialist technical searches or analysis of the environment in which the AI system operated.
• Consider where the evidence sits: important information may be held by the AI developer or provider rather than the organisation deploying the system. Understanding who controls relevant models, datasets, logs and documentation can therefore be an important early step when a dispute arises.
• Protect sensitive information: technical evidence may contain trade secrets, proprietary technology, commercially sensitive datasets or information about system vulnerabilities. Measures such as confidentiality clubs, redactions and controlled inspection arrangements may be needed to balance disclosure obligations against confidentiality and cybersecurity concerns.

The approach will also differ between jurisdictions. The English High Court has tools including extended disclosure, non-party disclosure, and bespoke inspection arrangements that can be adapted to technically complex evidence. The CAT has experience dealing with high volumes of structured data and can use approaches including expert-led disclosure and statistical sampling.

France takes a different approach. There is no general disclosure process equivalent to that in England and Wales. Instead, parties generally need to identify the evidence they require and seek targeted measures from the court, including, where appropriate, pre-action measures under Article 145 of the French Code of Civil Procedure.

Expert evidence: what you need to know

Expert evidence is also likely to play an important role in AI disputes, helping courts understand both how an AI system operated and what conclusions can reliably be drawn from the available technical evidence.

Key considerations include:

• Identify the expertise required early: an AI dispute may require expertise in machine learning, statistics, cybersecurity, data engineering or a specific industry. A general knowledge of AI may not be enough to address the particular technical issues in dispute.
• Consider independence and conflicts: suitably qualified specialists may be relatively limited in number and may already have relationships with relevant companies or technologies. Identifying potential experts and carrying out conflict checks early could therefore be important.
• Make complex evidence understandable: experts will need to translate highly technical material into terms the court can understand, explaining how a system operated, what the available evidence demonstrates and, equally importantly, what cannot reliably be concluded from it.
• Consider how expert evidence will be managed: English proceedings may involve party-appointed experts or, where appropriate, a single joint expert. Concurrent expert evidence, sometimes referred to as "hot tubbing", can also help narrow the areas of genuine technical disagreement. In France, courts generally appoint a neutral expert to conduct an adversarial expert process.

Preparing for an AI dispute

The evidential challenges created by AI make early preparation particularly important. Organisations should consider:

• preserving relevant models, prompts, outputs, logs, configurations and datasets as soon as a dispute is reasonably anticipated;
• mapping the developers, providers, deployers, users and data owners that may hold relevant evidence;
• considering whether specialist collection, review or controlled inspection arrangements will be required;
• putting appropriate confidentiality and cybersecurity protections in place for sensitive technical material;
• identifying suitable experts, and potential conflicts, at an early stage; and
• allowing additional time and cost for technical review, testing, disclosure and expert analysis.

Taking these steps early can help organisations preserve the evidence needed to reconstruct how an AI system operated, understand where key information is held and reduce the risk of evidential issues becoming more difficult or costly once proceedings are underway.

The final session in our AI x Dispute Resolution Webinar Series 2026 will consider how organisations can future-proof themselves against AI-related disputes. If you would like to attend this session you can register here, or contact your usual Simmons & Simmons contact or our AI x DR team.