That matters because medico-legal reporting is not ordinary drafting. A report may influence settlement, litigation strategy, rehabilitation, damages, credibility and the court’s assessment of causation and prognosis. If AI is used carelessly, a small error in a summary or chronology can become a much larger error in the final opinion.
AI can assist with report preparation. It may help organise records, identify dates, remove duplication and produce a first draft of a chronology. Used carefully, it may reduce administrative burden. But used badly, it can introduce false confidence into a process that already depends heavily on accuracy.
A hallucination does not have to be dramatic.
When lawyers talk about AI hallucinations, the obvious example is invented case law. That is serious, but medico-legal hallucinations may be quieter and harder to spot.
An AI tool may state that a claimant first reported neck pain two days after an accident when the record refers to back pain. It may describe medication history as an escalation when the prescription was only a repeat issue. It may state that there was no pre-accident psychiatric history because it failed to identify earlier entries for anxiety or antidepressant medication. It may treat a historic symptom as current, or a copied hospital letter as a new consultation.
These are not cosmetic mistakes. They can alter causation, baseline, prognosis and disability analysis.
In a whiplash claim, the timing of first symptoms may matter. In a psychiatric injury claim, early records may be central to whether symptoms were accident related. In a chronic pain claim, medication history and treatment response may affect credibility and prognosis. In a pre-existing condition case, one missed entry may change the entire baseline.
The risk, therefore, is not just that AI might invent something obvious. It is that it may quietly distort the evidence.
One of the most dangerous features of AI-generated text is its fluency. A poor human draft often looks poor, but a poor AI draft may look professional, structured and confident. That can make the error harder to detect.
This is especially risky in medico-legal reports because the writing may sound more certain than the evidence allows. An AI-generated paragraph may state that symptoms are consistent with the accident, that recovery is likely within a particular period, or that the records support the claimant’s account, without properly explaining the basis for those statements.
The expert must not confuse neat drafting with sound reasoning.
A report is credible because the opinion is supported by records, examination findings, chronology and clinical judgement. It is not credible simply because the prose is smooth. If AI makes a weak opinion sound stronger, it has not improved the report. It has made the weakness less visible, that is the real risk.
Chronology is particularly vulnerable. AI is likely to be used increasingly for chronology building because medical records are long, repetitive and poorly ordered. That is understandable. Chronology work is time-consuming, and tools that can order records by date or identify repeated entries may be useful.
However, a medico-legal chronology is not a neutral administrative list. It is often the foundation of causation.
The chronology tells the expert what happened before the accident, what changed afterwards, when symptoms were first reported, whether treatment escalated, whether recovery occurred and whether the claimant’s account fits the documents. If AI gets that sequence wrong, the report may start from the wrong factual basis.
This can happen in subtle ways. A tool may include duplicated entries as though they were separate consultations. It may miss handwritten or scanned material. It may misunderstand abbreviations. It may confuse the date of a letter with the date of the appointment. It may place a copied past medical history entry into the wrong part of the timeline.
For that reason, the expert should treat any AI-generated chronology as a draft aid, not as evidence. It must be checked against the source records before it is relied upon.
False sources are not only a legal problem, but the legal profession has also already seen the damage caused by false AI-generated citations. Medico-legal reporting has its own version of the same problem.
An AI tool may refer to guidance, studies, diagnostic criteria, recovery periods or treatment recommendations in a way that sounds authoritative but is inaccurate or unsupported. It may cite a guideline that does not say what the tool claims and it may invent a statistic about recovery. It may overstate what a medical paper proves, it also may produce a general proposition that is true in broad terms but wrong for the specific claimant.
This is especially problematic where reports discuss psychiatric injury, chronic pain, mild traumatic brain injury or prognosis. These are areas where confident but oversimplified statements can easily become misleading.
An expert should not include any authority, statistic, guideline or clinical proposition unless they can verify it. If the expert cannot identify the source and check the wording, it should not appear in the report. A false source in an expert report is not a harmless drafting issue, and it could also mislead the parties and the court.
AI may assist with administration, but the opinion must remain the expert’s own. That distinction is essential. An expert report is not simply a document that needs to be produced. It is professional evidence. The expert must understand the records, assess the claimant, consider material facts, address evidence that may detract from the opinion and explain the reasoning.
AI cannot assume that duty.
If an AI tool produces a summary, the expert must verify it. If it drafts a causation section, the expert must question the reasoning. If it suggests a prognosis, the expert must decide whether that prognosis fits the claimant’s actual recovery course. If it flags an inconsistency, the expert must decide whether that inconsistency matters medically. The signature on the report is not symbolic. It means the expert owns the opinion, and an expert should never be in the position of defending a conclusion because the system produced it. That answer will not assist the court, and it will not protect the report.
Hallucination is not the only danger. The same tools that may produce inaccurate output may also create confidentiality and data protection risks if used without proper safeguards.
Medico-legal material often includes sensitive medical records, psychiatric history, medication details, employment information, family circumstances and litigation documents. If that material is entered into an unsuitable AI system, the risk is not only that the output may be wrong. The risk is also that confidential information may be processed, retained or used in ways the expert does not understand.
Removing a claimant’s name may not be enough. Dates, occupation, accident circumstances, treatment history and personal details may still identify the individual when combined.
Before using AI with case material, the expert or organisation must know what data is being entered, where it is processed, whether it is retained, who can access it, whether it is used for training and what contractual safeguards exist.
If those questions cannot be answered, confidential case material should not be used with that tool.
One hallucination can contaminate several sections of a report. A wrong chronology entry may affect causation. An incorrect medication summary may affect severity. A missed pre-accident symptom may affect baseline. A false statement about treatment response may affect prognosis. Once the error appears in a draft, it may be repeated in the diagnosis, opinion, prognosis and answers to questions.
This is why AI output needs active checking, not passive reading.
The expert should check the points that matter most: symptom onset, pre-accident history, medication, treatment chronology, imaging, employment function, psychological symptoms, recovery pattern and any inconsistency relied upon. These are the facts that usually carry medico-legal weight.
Not every line of a long bundle will be equally important. However, every fact used to support the opinion must be traceable to the records, if it cannot be traced, it should not be relied upon.
AI becomes dangerous when it changes the expert’s behaviour. If it encourages the expert to read less of the source material, check fewer records or accept a chronology because it looks complete, then the tool has weakened the report. If it helps the expert locate relevant entries more efficiently and then verify them properly, it may improve the process.
The difference is professional discipline, the expert should approach AI output in the same way they would approach a draft prepared by someone else: useful, but not final; helpful, but not authoritative; capable of saving time but still requiring review.
The more central the output is to the opinion, the greater the need for verification.
Grammar assistance is one thing. AI-generated causation analysis is another. A draft chronology sits somewhere in between, but it still requires checking because chronology often drives the opinion.
Safe use should be clear, documented and, frankly, boring.
The organisation should approve the tools being used. Staff and experts should understand what can and cannot be entered. Confidentiality rules should be explicit. There should be a process for checking AI-generated summaries and chronologies against the source material. Any material reliance on AI should be capable of explanation.
The expert should also keep control of the final report. They should know which records were reviewed, what assumptions were made, what limitations remain and which parts of the opinion depend on incomplete evidence. If AI has been used to assist with a chronology or summary, the expert should be able to say that the output was checked and that the final opinion is their own.
AI should not be banned from medico-legal work simply because it can make mistakes. Human reviewers also make mistakes. The issue is whether the system has safeguards that catch errors before they enter the evidence.
A hallucinated fact should never survive into the final report because it sounded plausible. A source should never be cited because AI supplied it. A prognosis should never be accepted because it appeared in a draft. Every material point must be checked against the evidence and clinical reasoning.
That is where expert judgement remains irreplaceable.
AI can assist the expert with the burden of information. It cannot decide what the information means.
Medico-legal evidence depends on trust. The court trusts that the expert has reviewed the material properly. The parties trust that the report is independent and accurate. The claimant trusts that confidential information will be carefully managed. The expert’s signature represents all of that.
AI hallucinations threaten that trust because they can make false material look dependable.
The answer is not panic, and it is not blind adoption. The answer is controlled use, clear accountability and careful verification. AI may help organise records, improve drafting and reduce delay, but it must never become the hidden author of unverified opinion and the expert must remain able to explain every material conclusion in the report.
Where the evidence is uncertain, the report should say so. Where the records are incomplete, the report should say so. Where AI output has assisted, it must be checked. Where it cannot be checked, it should not be used.
In medico-legal evidence, a hallucination is not just a technical error.
It is a risk to the integrity of the opinion.
