The trouble with AI? It keeps making things up
Large language models are good at lying – and good at disguising their lies
Daily Telegraph 13 December 2025 Referencing added.
- “There is no hierarchy of rights; all are to be treated with equal respect.” What an unobjectionable phrase that is. So reasonable-sounding, so plausible, that the eye simply glides over it. Which is precisely what should make the eye suspicious.
- Just two months ago, the Courts and Tribunals Judiciary issued new guidance on how to stop AI-generated guff sneaking into legal documents. added link AI cannot replace direct judicial engagement with evidence and legal argument. Now, the already fraught and prolonged tribunal case between NHS Fife and Sandie Peggie (the nurse who complained after having to share a changing room with a transgender doctor) added link has been thrown into disarray by the discovery of “false and non-existent quotes” in the judge’s 300-page ruling. AI does not show - long page. more Beth Upton
- Two passages, both presented as quotations from previous tribunal rulings, have so far been identified as fabrications. The most likely explanation seems to be that someone used AI to help put together the judge’s briefings. As a result, the entire ruling may be withdrawn or overturned on appeal. All that effort, expertise, commitment and stress, made redundant by the hyperactive imaginings of a robot.
- It used to be thought that LLMs (large language models) would never be a match for human writers because algorithms don’t have an imagination. If only! Turns out, AI is brilliant at making things up. That’s literally what it does, word by word. LLMs use statistical probability, based on analysis of all the billions of sentences on the internet, to work out which word should come next. This produces the kind of smooth forward propulsion, one word inexorably leading to another, that most writers can only dream of.
- But it also produces reams of fiction where facts are required. And its nonsense is hard to spot, because the mathematically-calculated placing of each word makes it frictionless. There is nothing for the eye to snag on. This makes it doubly-dangerous as a research tool. It’s good at lying, and good at disguising its lies.
- My children have, reluctantly, given up on asking ChatGPT to do their homework for them. It wasn’t worth the risk or bother in the end. By the time you’ve meticulously combed through its silky prose, picking out the inaccuracies like particularly evasive nits, and then roughing up whatever remains so that it reads like something a 21st century teenager might actually write – well, what’s the point? It would be quicker to do the assignment from scratch.
- I had hoped to use AI to do much of the drudge work of translation and transcription for a book I am working on. (A history of reviving the dead, which requires a lot of archival research as well as recorded interviews.) But it keeps serving me dollops of pure fantasy. It made up, out of nowhere, a medieval miracle that I ardently wish had been true, in which a non-existent Welsh saint succeeded in re-attaching the head of a decapitated boy. Being unable to cope with regional accents, it transcribes all northern or Scottish interviews into a series of childlike, disembodied exclamations.
- Recently, I asked it to transcribe a document for me, using a photograph of a slightly dog-eared sheet of paper. It produced what appeared to be a fluent stream of prose, but stuffed full of invented quotations. On closer inspection I realised that, every time the robot had encountered a shadowy patch of crumpled paper, it had simply made up what it couldn’t read.
- It was, to be fair, very apologetic – and honest about its failings. “I cannot, with my current capabilities, give you a mathematically guaranteed zero-hallucination transcription,” it confessed. “I’m biased to fill a gap rather than leave it blank.”
- I couldn’t help feeling a small, pathetically enjoyable rush of vindication. All writers love to see AI mess up, because it’s going to wipe out much of our profession. We’re like 18th century weavers, bitching about the poor quality yarn produced by the spinning Jenny.
- AI will win, because everything frictionless is irresistible to consumers. But we can already see how the truth will be lost.