is a freelance journalist who writes for Philonomist, Philosophie Magazine, and Chut!, a French magazine dedicated to digital technology. His work focuses on new technologies and the ways in which they shape us.
Picture © Author’s personal collection
Artificial intelligence is man-eating, but not cannibalistic, according to a recent study. If, instead of training with human-made data, it digests content generated by another AI, it inevitably suffers a spectacular degradation of its performance. And while the internet is flooded by artificial content and human data becomes more scarce, this glitch could compromise AI’s progress.
AI cannot bear to devour its own creations. This is the strange but critical conclusion reached by six researchers from the universities of Oxford, Cambridge, London, and Toronto, in an article from May 2023. To understand, we need to keep in mind that language models like GPT-4 or image generation software like Stable Diffusion base their performance on a machine learning process which requires considerable amounts of data. To generate a decent image of a cat, for example, Stable Diffusion must first be able to see lots of real cats, of different colours and sizes, in infinitely varied positions, in order to then predict what we mean by the word.
The machine’s training ground is the internet itself, which is full of such data. Except that, since the launch of the remarkably popular ChatGPT in November 2022, everyone has been using these powerful tools to generate artificial texts and images which have in turn been flooding the internet. In other words, the pantry that allows AI to feed on data of human origin is gradually being polluted by their own emanations.
Stuck in an AI loop
So what happens when an AI absorbs the fruit of its own statistical flesh? The result, say the authors of this study, is worse than indigestion. “When we use a model to train another model, it means basically the imperfections of the first model are learned by the second model while having its own imperfections. Imperfections on top of imperfections,” Oxford researcher Ilya Shumailov explains over the phone. “What we…
Subscribe to Philonomist and gain free access to all our content and archives for 7 days. You'll also receive our weekly newsletter. No commitment. No bank details required.
You're already subscribed to Philonomist via your employer?
Connect to your account by filling in the following details (please provide your professional email address).
