Historian Tom Holland was guest-editing the ‘Today’ programme on BBC Radio 4 recently. He spoke about his experience of AI ‘hallucinations’, that now increasingly well-known phenomenon whereby Artifical Intelligence makes up information in response to a question.

Tom Holland had been trying to ascertain whether the ancient Greek poet Pindar had mentioned a certain thing or not (I forget the details). Holland thought not, but he decided to check with AI. Lo and behold the answer was that yes, Pindar did indeed mention this thing, and the poet’s lines were actually quoted. Tom Holland was taken aback, but went to some lengths to check the reference, and could not find it – not even in the British Library. Finally he realised that the answer given by AI must be an example of ‘a hallucination’. I am not sure if anyone quite understands why these hallucinations occur. ‘Large language models’ only contain what is put into them, but it seems that AI can sometimes conjure new information from old.

I myself first came across AI hallucinations when researching the life of the Irish composer John Field. John Field had a painful illness towards the end of his life. I wondered why modern historians could confidently assert that the illness was rectal cancer. It seemed unlikely that such a diagnosis was publicly known in the 1820s and 30s. So I asked AI if it could tell me why music historians thought that John Field had suffered from rectal cancer. It replied that Field’s illness was discussed in certain letters exchanged between Field and the pianist-composer Johann Baptist Cramer. This seemed plausible, as the two men knew one another. Both had studied with Clementi, and both were pallbearers at Clementi’s funeral.

But I had never heard of these letters, so I asked if AI could give me the sources. It replied evasively that the letters were mentioned here and there in the work of other writers. So I asked where I could find the references.

Eventually AI started backtracking. ‘I apologise for having said that there were letters between John Field and Johann Baptist Cramer in which rectal cancer is mentioned. There are no such letters.’

For me, the apology was as surprising as the original information. If there were no such letters, why did AI state that there were? I had supposed that it merely combed through its archive of information on a given topic and summarised it at lightning speed. But … inventing a correspondence which didn’t exist?

I suppose that by querying its answers and forcing it to the point of ‘apologising’, I did my own little bit of AI training. Because today, when I repeated my original question, it did not claim there were any particular letters on the subject, but confined itself to an overview of Field’s journeys to seek medical treatment.

As Tom Holland pointed out, if researchers innocently take these ‘hallucinations’ on trust, the errors could work their way into books and articles and thus become embedded in the literature. Quite a scary prospect when you think about on the larger scale!