Monday 9 November, 2026 – by John Naughton – Memex 1.1
John Naughton in trenchant mode:
“ Here’s an experiment you probably can’t try at home. Take a large language model (LLM) and prompt it with a passage about English church architecture of the 14th century; you know the kind of stuff – towers, master masons, the perpendicular style, and so on. Straight out of Pevsner, if you like. Ask it to carry on writing. Then train a fresh model on what the first one wrote, and ask it to do the same.
Keep going, generation after generation. By the ninth, you find that the machine has given up talking about churches and is producing an earnest list of jackrabbits with tails of various colours: black, white, blue, red and yellow.
You think I made this up? Nope. It’s one of the illustrations in a paper published in Nature by researchers at, among others, Oxford, Cambridge, Imperial College London and Toronto University who wanted to know what happens when AI systems start learning from the work of other AI systems. Their answer: the machines go gently, and then not so gently, potty. The technical term for it is “model collapse”.
…
Unless the model collapse problem is solved, we’re heading into what Jathan Sadowski, a technology researcher at Melbourne-based Monash University, memorably called “Habsburg AI”. He used it to describe models trained so heavily on the output of other LLMs that they become inbred mutants, rather like the Spanish bloodline of the famous European dynasty, which inbred itself to extinction. Or, to put it more crudely, it may mean a kind of accidental enshittification of LLMs.”
Well, after all, Gregor Mendel was firmly based within the Habsburg empire.