AI models collapse when trained on recursively generated data

Researchers uncover 'model collapse' when AI is trained on data generated recursively by prior models. The study, led by Ilia Shumailov and Zakhar Shumaylov, details how indiscriminate use of model-generated content leads to fundamental defects in evolving models.

• Article published on 24 July 2024
• Stable diffusion revolutionized image creation
• GPT-n models susceptible to 'model collapse'
• Real human interaction data increasingly valuable
• Access to original data distribution is crucial