AI and Expert Evidence: Useful Tool, Dangerous Shortcut

Artificial intelligence is now easy to access, quick to use, and often superficially impressive. It can produce polished text in seconds. It can summarise, organise, proofread, draft, and present information in a way that looks coherent and authoritative.

It can also produce plausible nonsense.

That is the central problem for expert witnesses. AI can be helpful, but it is not reliable simply because it is fluent. It does not relieve the expert of responsibility. It cannot assume the expert’s duty to the court. It cannot be permitted to form the expert’s opinion. It cannot be treated as a substitute for professional judgement.

For expert witnesses, the starting point is not enthusiasm about technology, nor anxiety about it. The starting point is the expert’s duty.

The expert’s duty comes first

The expert who signs a report is responsible for the whole of that report. That remains true regardless of the tools used in preparing it.

An expert may use secretarial assistance, dictation software, spellcheckers, indexing tools, medical records software, literature searches, legal databases, administrative support, junior researchers, or AI-assisted systems. None of those tools transfers responsibility away from the expert.

The expert’s duty to the court is personal and non-delegable.

That duty cannot be delegated to AI.

Experts in civil proceedings are governed by the Civil Procedure Rules, the Practice Direction, and the specific requirements for expert evidence. Those are not optional recommendations or informal best-practice tips. They are the rules that determine whether an expert report is compliant.

This point is sometimes overlooked. An expert does not merely need to produce a report that is persuasive, well-written, or useful to the instructing party. The report must comply with the procedural requirements that govern expert evidence. If it does not, it may fail in its purpose.

That is true across expert disciplines. A psychiatrist, hand surgeon, engineer, accountant, toxicologist, architect, housing expert, or materials scientist may work in very different professional worlds, but the court-facing duty is the same. The expert must provide independent, impartial, properly reasoned evidence within their expertise.

AI does not change that duty. It simply introduces a new category of tool, with new risks.

Is the use of AI by experts permissible?

In principle, yes.

The use of AI by expert witnesses is not inherently impermissible. There is no sensible basis for saying that an expert must never use AI for any purpose at all.

Indeed, it may eventually become difficult to argue that experts should never use AI. If a tool helps an expert produce a more accurate, complete, internally consistent, and reliable report, then there may be circumstances in which not using that tool could itself be criticised.

That does not mean that experts should embrace AI uncritically. It means that the question is more nuanced than simply “AI: yes or no?”

The real questions are:

• What was AI used for?
• What information was given to it?
• Was the system secure?
• Was confidential or privileged material protected?
• Was the output checked?
• Did the AI influence the expert’s opinion?
• Was the use recorded?
• Was the use disclosed, or at least capable of being disclosed?
• Could the expert explain the process if asked in court?

Those are the questions that matter.

The developing guidance

Guidance on AI and legal work has developed quickly. It is more developed for judges, solicitors, and barristers than it is for expert witnesses, but it is still useful for experts to understand the direction of travel.

The sources of guidance and commentary currently include:

• the Civil Procedure Rules and Practice Direction 35;
• the Criminal Procedure Rules, where relevant;
• guidance for judicial office holders on the use of AI;
• the Solicitors Regulation Authority’s guidance and compliance material;
• the Bar Council’s guidance and considerations on AI use;
• guidance and commentary from the Academy of Experts;
• guidance from medico-legal professional organisations, including the Medico-Legal Knowledge Partnership;
• consultation work by the Civil Justice Council;
• reported cases in which AI misuse has caused difficulties in litigation.

There are broadly two approaches in this emerging material.

The first approach says that existing principles are sufficient. Lawyers, judges, and experts already have duties of competence, confidentiality, independence, accuracy, and candour. Those duties apply regardless of whether the tool being used is a textbook, a search engine, a legal database, a spreadsheet, a dictation system, or a generative AI model.

The second approach says that AI creates sufficiently distinctive risks that specific guidance is required.

In practice, both positions have merit.

The existing principles remain fundamental. An expert’s duties do not change simply because a new tool is available. But AI also creates practical risks that are sufficiently novel, or sufficiently easy to misunderstand, that more specific guidance is needed.

That is particularly true in relation to confidentiality, hallucinated authorities, fabricated citations, data retention, disclosure, and the extent to which AI has influenced the content of a report.

For expert witnesses, the position remains less settled than one might wish. There is helpful non-binding guidance, but there remains uncertainty about exactly what must be disclosed, how AI use should be recorded, and what uses will be regarded as acceptable in future litigation.

That uncertainty should make experts cautious.

AI is not a textbook, database, or reference engine

One of the most important things to understand about generative AI is that fluent output is not the same as reliable output.

At root, large language models generate statistically likely text based on patterns in training data. They can produce answers that appear polished, coherent, and confident. But they do not verify truth in the way an expert must. They are not textbooks. They are not legal research databases. They are not medical literature databases. They are not inherently reliable citation engines.

When AI produces false information in a fluent and confident form, this is often described as a hallucination.

That is not a minor technical problem. In professional work, it can be very serious.

AI may produce:

• non-existent legal cases;
• real case names attached to false propositions;
• inaccurate summaries of judgments;
• fabricated journal articles;
• plausible but false quotations;
• incorrect references;
• misleading summaries of clinical guidance;
• confident but unsupported factual assertions.

For expert witnesses, the risk is obvious. An AI system may generate an apparently impressive list of journal articles or textbook references. Some may be real. Some may be inaccurate. Some may not exist. Some may exist but not support the proposition for which they are cited.

That creates a direct risk to the expert’s credibility.

An expert cannot rely on a source simply because AI has produced it. If AI provides a case, check the case. If it provides a journal article, check the article. If it gives a quotation, check the original source. If it summarises guidance, read the guidance yourself.

The rule is simple: verify everything.

Lessons from reported cases

The legal cases in which AI misuse has already caused difficulty are important because they show what can go wrong.

In some cases, lawyers have put forward authorities that were generated or distorted by AI. The problem is not merely that the cases were unhelpful. The problem is that some were fabricated or materially inaccurate.

In the Ayinde case, legal material was advanced which included fabricated authorities. In the Al-Haroun case, a number of false or fabricated cases were relied upon in legal submissions. These cases illustrate the danger of using AI as if it were a reliable legal research tool.

The Munir case illustrates an additional danger: confidentiality and privilege. Confidential and legally privileged client material was reportedly put into a public AI system. That raised serious issues about the handling of sensitive information, the protection of privilege, and the professional responsibilities of the solicitor involved.

For expert witnesses, the lesson is direct.

If a solicitor cannot excuse the use of fabricated AI-generated authorities, an expert cannot excuse fabricated AI-generated research. If a solicitor cannot avoid responsibility by blaming a junior member of staff or a tool, an expert cannot avoid responsibility by blaming AI.

The person who signs the document bears responsibility for it.

Confidentiality is a central risk

For experts, the most immediate risk may not be hallucination. It may be confidentiality.

Expert witnesses routinely handle highly sensitive material. Depending on the case, that may include:

• medical records;
• psychiatric histories;
• psychological assessments;
• social care records;
• educational records;
• employment records;
• criminal allegations;
• witness statements;
• photographs;
• imaging reports;
• correspondence with solicitors;
• privileged material;
• intimate personal information;
• third-party information.

This material should not be uploaded into public AI systems.

By “public AI”, I mean systems where the expert cannot be confident that the material will remain confidential, will not be retained beyond what is necessary, will not be reviewed, and will not be used for model training.

The problem is not solved merely because a system is popular, convenient, or produced by a well-known technology company.

Nor is it enough to assume that a paid version is automatically safe. The expert needs to understand what protections actually apply.

Important questions include:

• Is the data used for model training?
• Is the data retained?
• If so, for how long?
• Where is the data processed?
• Who can access it?
• Is the processing covered by an appropriate contractual arrangement?
• Does the system comply with relevant data protection obligations?
• Is the system appropriate for confidential medico-legal material?
• Has the instructing party approved or authorised its use?
• Could the expert explain the use if challenged?

Privacy settings and contractual terms may change. A system that appears safe at one point may not remain safe indefinitely. Users may be presented with updated terms, revised policies, or new default settings. It is easy to click through such updates without appreciating their significance.

That is dangerous in expert work.

If confidential case material is uploaded into an inappropriate AI system, the expert may have created a serious problem. It may be a confidentiality problem, a data protection problem, a privilege problem, a professional conduct problem, or all of those things.

The safest practical rule is straightforward:

Do not upload confidential, privileged, identifiable, or case-specific material into public AI systems.

Cloud processing and due diligence

Some AI systems process data in the cloud. That is not automatically improper. Many professional systems now operate in cloud environments, including systems used by law firms, hospitals, insurers, and public bodies.

The issue is not whether something is “cloud” or “not cloud”. The issue is whether the expert understands the data pathway and whether the safeguards are adequate.

If an expert uses any system to process case material, particularly medical records or other sensitive documents, the expert should understand:

• where the data goes;
• who controls the system;
• whether third-party processors are involved;
• whether data is retained;
• whether data is used to train models;
• whether the data can be accessed by humans;
• whether the system is compliant with applicable data protection requirements;
• whether there is an audit trail;
• whether the instructing solicitor knows and approves of the process.

This is not a criticism of cloud systems. It is a reminder that expert evidence often involves unusually sensitive material.

The expert must be able to justify how the material has been handled.

AI-generated chronologies and summaries

One area in which AI is already becoming relevant is the production of chronologies and summaries from large records bundles.

Many experts are familiar with cases involving thousands of pages of medical records, sometimes poorly ordered, duplicated, fragmented, or provided in awkward file formats. AI tools may assist in organising such material. They may identify dates, group events, remove duplication, summarise consultations, and produce draft chronologies.

That may be useful.

But a chronology is not neutral merely because it is chronological.

Choices are made about what to include, what to omit, how to describe an event, and whether something is relevant. If AI omits a key consultation, mislabels a diagnosis, misunderstands an abbreviation, or overstates the significance of an entry, the expert may be misled.

This matters because expert opinion is often built on the factual foundation created by the records review. If that foundation is wrong, the opinion may be weakened.

Experts should therefore be cautious about relying on AI-generated chronologies or summaries. They may be useful as an aid, but they are not a substitute for expert scrutiny.

If a chronology has been produced by someone else, or by an AI-assisted process, the expert should be clear about what has been checked and what has been relied upon.

The expert remains responsible for the report and for any opinion based on the records.

Disclosure and transparency

Disclosure is one of the most difficult areas.

Practice Direction 35 requires experts to give details of literature or other material relied on in making the report. If an expert relies on a published paper, a textbook, a guideline, a calculation, or specialist material, that may need to be identified.

AI complicates the position.

If AI has merely been used as a spellchecker, disclosure may not be necessary. If AI has been used to calculate, summarise, structure, analyse, draft, or check material that contributes meaningfully to the report, the position is more difficult.

The current rules may not provide a complete answer to every possible use of AI. That does not mean the issue can be ignored.

Experts should assume that they may later be asked about AI use. Reports often reach trial years after they are written. Guidance and expectations may have changed by then. A report written today may be scrutinised in court in three or four years’ time, by which point disclosure expectations may be clearer and more demanding.

The expert must be able to answer honestly.

A useful practical test is this:

If you would feel uncomfortable explaining your use of AI under cross-examination, do not use it in that way.

Even where disclosure is not presently mandatory, experts should keep a record of meaningful AI use. That record should include:

• the tool used;
• the date of use;
• the task for which it was used;
• the information provided to the tool;
• the output generated;
• whether the output was accepted, rejected, or amended;
• how the expert verified the output.

This may feel burdensome, but it is proportionate where AI use has contributed to the preparation of expert evidence.

Low-risk and higher-risk uses

Not all AI use carries the same level of risk.

At the lower-risk end are uses such as spelling, grammar, formatting, and typographical checking, provided confidential information is not uploaded into an insecure system.

AI may also be useful for generic, non-confidential sense-checking. For example, an expert might ask for a general list of differential diagnoses associated with a non-identifiable clinical presentation. The expert may then consider whether any of those possibilities are relevant.

That is not the same as asking AI to diagnose the claimant. It is not the same as asking AI to determine causation. It is not the same as allowing AI to form the opinion. It is closer to using a checklist, a textbook index, or a teaching prompt.

There are also useful quality-control applications. AI can be used to check a report for internal consistency, unanswered questions, typographical errors, missing sections, date discrepancies, or contradictory statements.

At the higher-risk end are uses such as:

• asking AI to analyse the evidence;
• asking AI to determine diagnosis;
• asking AI to advise on causation;
• asking AI to assess credibility;
• asking AI to evaluate vulnerability under cross-examination;
• asking AI to draft substantive opinion;
• relying on AI-generated citations;
• relying on AI summaries of records without checking them;
• uploading confidential material into public systems.

Those uses may compromise confidentiality, independence, reliability, or the expert’s ability to defend the report.

AI should not form the opinion

The expert’s opinion must be the expert’s own.

AI should not decide the diagnosis. It should not determine causation. It should not assess credibility. It should not determine prognosis. It should not decide whether symptoms are attributable to an index event. It should not determine whether a claimant meets a legal test. It should not supply the expert’s reasoning.

Those are expert functions.

An expert may consider many sources when forming an opinion: the records, the interview, the examination, the literature, the expert’s training, professional experience, clinical judgement, and the legal questions asked. AI should not be added to that list as if it were an independent source of expert judgement.

There is also a risk in allowing AI to suggest wording for substantive opinions. An expert who adopts AI-generated language may later struggle to explain why a particular phrase was used. In court, “the computer suggested it” is not an answer.

The expert must be able to own the opinion, explain the reasoning, and defend the wording.

AI as a proofreading and quality-control tool

One of the most promising uses of AI is in proofreading and consistency checking.

Human proofreaders often miss certain types of errors. We read what we expect to see. We may notice typographical errors in paragraphs but miss them in headings. We may fail to notice a missing “not”. We may overlook an inconsistency because we understand what was intended.

AI-assisted checking can help identify:

• inconsistent names;
• inconsistent dates;
• incorrect locations;
• missing negatives;
• typographical errors;
• contradictions between sections;
• unanswered questions from the letter of instruction;
• missing declarations;
• unsupported assertions;
• discrepancies between the summary and the body of the report;
• headings that do not match the content;
• references to the wrong party;
• internal inconsistencies in the clinical history.

This is a legitimate and useful distinction.

Using AI to identify a possible inconsistency is not the same as asking AI to form an opinion. It is a quality-control process. The expert remains responsible for deciding whether the flagged issue is real, whether it matters, and whether the report needs amendment.

In this role, AI may be particularly valuable because it can identify patterns that humans easily miss. It may detect a road name that does not match the stated location, a date that does not fit the chronology, or a section in which the claimant’s account appears to contradict an earlier section.

That does not make AI the expert. It makes it a tool.

The risk of professional embarrassment

Some mistakes in expert reports are minor. Others can seriously damage credibility.

Anyone who gives expert evidence understands the importance of confidence in the report. If cross-examination begins with obvious errors, inconsistencies, or unanswered questions, the expert is immediately placed in difficulty.

A typographical error may not matter. But an incorrect name, a contradictory history, a missing negative, a wrong date, or an unanswered question from the letter of instruction may matter a great deal.

The problem is not only the specific error. The problem is the inference that may be drawn from it.

If the expert has missed an obvious error, what else has been missed? If the report contains internal contradictions, how careful was the analysis? If the expert has not answered all the questions asked, how reliable is the opinion?

AI-assisted quality control may help reduce those risks. It should not be seen as a substitute for careful expert review, but it may provide an additional layer of scrutiny.

Practical principles for expert witnesses

The safest approach is to treat AI neither as a miracle nor as a menace.

It is a tool. Some uses are sensible. Some are dangerous. Some may be acceptable only with proper safeguards. Some should be avoided.

The following principles are a practical starting point.

1. The duty to the court comes first

The expert’s duty is personal and non-delegable. AI cannot reduce, dilute, or transfer that duty.

2. Do not use public AI for confidential case material

Confidential, privileged, identifiable, or case-specific material must never be uploaded into public AI systems.

3. Understand the system being used

Experts should understand whether data is retained, whether it is used for training, where it is processed, who can access it, and what contractual safeguards apply.

4. Do not let AI form or influence the expert opinion

AI may assist with organisation, proofreading, and checking. It must not decide the opinion.

5. Verify everything

Cases, journal articles, quotations, clinical guidelines, factual propositions, and calculations must be checked against reliable sources.

6. Beware fluent nonsense

AI output can be polished, confident, and wrong. Fluency is not reliability.

7. Be cautious with AI-generated chronologies and summaries

They may be useful, but they may omit, distort, or misclassify important material.

8. Keep a record

Experts should keep a clear record of meaningful AI use, including the tool used, the information provided, the output generated, and how that output was handled.

9. Be ready to disclose

Even if disclosure is not currently required for every use, experts should assume they may later be asked about AI use and should be ready to explain it.

10. Apply the cross-examination test

If you would feel uncomfortable explaining your use of AI in court, do not use it in that way.

Seven rules for safe AI use

A shorter version of those principles is as follows:

1. The duty to the court comes first.
2. Never upload confidential case material into public AI systems.
3. Do not let AI form or influence the expert opinion.
4. Verify everything, especially cases, journal articles, quotations, and citations.
5. Beware fluent and compelling nonsense.
6. Keep a record of meaningful AI use.
7. Disclose, or be ready to disclose, that use.

Conclusion

AI is likely to become increasingly common in professional and expert work. It may help experts produce reports that are more accurate, more complete, and more internally consistent. It may help identify errors that human readers miss. It may assist with proofreading, formatting, organisation, and quality control.

But it also creates serious risks.

It can generate false information. It can fabricate references. It can misstate the law. It can summarise inaccurately. It can create confidentiality and privilege problems. It can tempt experts into accepting language or reasoning that is not truly their own.

For expert witnesses, the correct starting point is the expert’s duty to the court.

If AI helps the expert fulfil that duty, it may have a legitimate role. If it compromises that duty, it should not be used.

The key question is not simply, “Can AI do this?”

The better question is:

Can I still properly own, explain, and defend the opinion I have signed?