Yes, an expert witness can use AI, for a defined part of the work. AI is suited to the reading: extracting facts from a medico-legal bundle, building a first chronology, finding every entry that mentions a symptom or a drug. The opinion is a different matter. It must be the expert's own, every AI-produced fact must be verified against its source page, and the Part 35 duty to the court is untouched by the tooling. This post works through what that means in practice.

What does Part 35 actually require?

Three requirements carry the weight. The expert's overriding duty is to help the court on matters within their expertise, and that duty overrides any obligation to the instructing party. The report must be the expert's independent product, uninfluenced by the pressures of litigation, stating the substance of the instructions and the basis of each opinion. And the report ends with a statement of truth: the expert confirms they understand their duty, that the opinions are their own, and that the facts stated are true.

None of this mentions software, and none of it needs to. The rules are written around responsibility rather than method, which is exactly why they handle AI without amendment. Whatever produced a sentence, the expert who signs the report owns it.

Does using AI change who is responsible for the opinion?

No. The position is the same as it has always been for any assistance an expert uses, whether a paralegal's document sort, a textbook, or a search across the bundle. The expert may take help with the reading, and remains answerable for everything the report asserts. AI compresses days of reading into hours, and that changes the economics of an instruction. It changes nothing about authorship.

The statement of truth draws no distinction between sentences the expert typed and sentences the expert accepted. Every one of them is the expert's own.

The practical consequence is a simple discipline: an AI-surfaced fact enters the report only after the expert has opened the cited page and confirmed it. A tool that attaches the source page to every fact makes that discipline fast. A tool that returns confident, uncited answers makes it impossible, and an expert relying on one is signing for statements they cannot trace.

Should an expert disclose the use of AI?

The Civil Procedure Rules do not name AI, so there is no express disclosure rule to point to. The existing duties still reach the question from two directions. Practice Direction 35 requires the report to say what the opinions are based on and which materials were relied on, and cross-examination can probe how any part of the report came to exist. An expert who cannot explain their own method is in difficulty whether or not AI was involved.

The judiciary has meanwhile published its own guidance on AI for judicial office holders, which signals how courts are thinking: AI may assist with summarising and administrative work, its output must be checked, and responsibility stays with the person who signs. The safe assumption is that the same logic will be applied to experts. Record what the tool did, keep the verification trail, and be ready to explain both.

What can AI safely do in report preparation?

The safe territory is the work that is checkable against the record:

  • Data extraction and indexing. Typed facts such as diagnoses, medications, procedures and admission dates, pulled from every page into tables the expert can sort and verify.
  • Chronology building. A dated first-pass sequence of events, each entry anchored to the document and page it came from.
  • Targeted questions. Whether a drug was ever prescribed, when a symptom was first recorded, with the citation attached to the answer.
  • Summarisation for orientation. An overview of the bundle before the close reading starts, with gaps and duplicates flagged.

The boundary sits where checkability ends. Whether a standard of care was breached, whether causation is made out, how to weigh conflicting entries: these are the questions the expert is instructed to answer, and a machine's view on them has no status in the report. The division of labour is covered in more depth in AI for medical record review: where it helps and where judgement stays, and the tooling side on the AI medical record review page.

What are the risks when AI is used badly?

Two failure modes matter most. The first is the unverified assertion. Language models answer fluently when they should say the record does not support an answer, and courts in several jurisdictions have already criticised filings built on fabricated material. An expert whose report inherits an invented fact has a professional problem the statement of truth makes personal.

The second is data protection. Medical records are special category data, and the ICO's guidance on AI and data protection applies in full. Uploading a claimant's records to a consumer chatbot hands them to an external service on that service's terms. The governance questions come before the clever features: where does the data go, who can see it, and is every step logged.

How should an expert document AI use?

A short, contemporaneous record answers most future questions before they are asked: which tool and version was used, what it was used for, what it produced, and how the output was verified. If the tool attaches source pages, the verification record largely writes itself, because every fact that entered the report carries the page the expert checked it against. If the instructing solicitor asks whether AI was used, the answer is then a paragraph rather than a problem.

Why it matters

The rules already draw the line correctly: help with the reading is legitimate, and the judgement belongs to the expert. That is the principle ALLDOQ Verify is built on. It reads the whole record, extracts typed facts, builds a chronology and answers questions with the source page attached, on infrastructure the organisation controls, with every step logged for audit. The expert gets the bundle read in hours instead of days, and signs a report in which every fact can be traced to the page it came from, which is what Part 35 has required all along.

Sources