AI allegations against experts have reached the tribunal room, and one has now failed in the open. In LR (Male VOTs) Albania CG [2026] UKUT 00315 (IAC), issued on 13 August 2026, counsel for the Home Office put it to a country expert that his report was not his own work but the product of generative AI. The panel of three sitting at Field House found no proper foundation for so serious an allegation. It then took the three signs relied on and explained why none of them carried the point.
The case was country guidance on Albanian male victims of trafficking, heard over nine days between March and July this year. Professor Dalakoglou answered 199 written questions from the Home Office, his replies running to a further sixty-nine pages, and he then gave oral evidence for a whole day. Part of that day went on the suggestion that a machine had written what he had signed.
The three signs said to give it away
The allegation rested on three things. The report repeated the same phrases in places. It failed to mention the 2025 Trafficking in Persons report. And it used the expression to my opinion.
The tribunal took each in turn. The repetition traced back to the questions themselves, which overlapped, because the professor had answered what the appellant's solicitors put to him. The other two pointed the opposite way once the panel thought about how these tools behave, since a machine would have been more likely to cite recent material and to use proper English. The missing report bore on the weight the tribunal gave his evidence, which is a separate question altogether, and the phrase in dispute belonged to an academic writing in his second language who went on using it while he gave oral evidence.
Why an unfounded allegation costs more than a bad point
The panel called the suggestion a serious allegation and found no proper foundation for it. That wording repays attention. An advocate who says a report is thin is making an ordinary submission about quality. An advocate who says a machine wrote it is saying the expert signed work that was not his, which is a charge about honesty. A tribunal that rejects the first moves on without comment. A tribunal that rejects the second records who made it and why it failed.
The same cross-examination did land a point the panel treated as its most serious concern, about gaps in the identification of male victims of trafficking. That point had to make its way past two allegations the panel had already dismissed as unfounded. Weight came off the professor's evidence in the end, for reasons that had nothing to do with AI.
What would carry the point
Nothing in the judgment says an expert report can never be challenged on this ground. It says these three things did not do it. Repetition, an overlooked source and a phrase that reads oddly are the marks of a person writing at length under pressure as readily as the marks of a machine.
What would carry the point is harder to gather and simpler to state. Fabricated or miscited authorities. Sources in a bibliography that turn out not to exist. Assertions of fact that trace back to nothing. Passages that contradict what the expert says elsewhere in the same document. A tribunal can test each of those against the record in front of it. It cannot test a feeling about prose style.
What your firm should do with this
If you instruct experts, raise AI at instruction rather than at the hearing. A short line in the letter of instruction, asking the expert to confirm whether generative AI played any part in preparing the report and in what way, gives you an answer on paper long before anyone asks for one in public. Most experts will say it played none. Some will say they used a tool to search for material or to tidy their English, and that is an answer worth holding early rather than hearing for the first time in cross-examination.
If you are on the other side and you suspect a report was machine written, test the citations and the sources before you put anything to the witness. Where the sources check out, the point is not there to be made, whatever the prose reads like. Putting it anyway invites the finding this panel made, and it spends credit you will want for the points that land.
The judgment is published as LR (Male VOTs) Albania CG [2026] UKUT 00315 (IAC) on Find Case Law, which is open to read and asks nothing of you. The passage on the AI allegation begins at paragraph 157.
If your firm has no settled line on asking experts about AI, or on answering when you are asked, we work that through in an afternoon: start with a conversation.