While many advances in LLMs have been hailed as success stories – for instance the ability of Google’s Gemini to generate texts and images – these models have been trained on human-produced content. When such models instead begin to use data that has itself been generated using Artificial Intelligence, quality is compromised. Indeed it can, ultimately, lead to 'model collapse':

'Model collapse is a degenerative process affecting generations of learned generative models, in which the data they generate end up polluting the training set of the next generation.Being trained on polluted data, they then mis-perceive reality.'

Shumailov, I., Shumaylov, Z., Zhao, Y. et al. 'AI models collapse when trained on recursively generated data', Nature 631, 755–759 (2024). 

The article is available here.

Professor Yarin Gal is Tutor in Computer Science at Christ Church and Dr Ilia Shumailov was, until recently, a Junior Research Fellow at the college. The article was written in collaboration with PhD student, Zakhar Shumaylov, and Professor Ross Anderson – both of the University of Cambridge – along with Dr Yiren Zhao, Imperial College, London, and Professor Nicolas Papernot, University of Toronto.