TL;DR
A generative AI failure mode where models trained recursively on synthetic data progressively lose quality, diversity, and accuracy over successive generations.
Model collapse occurs when a neural network is repeatedly trained on datasets generated by prior AI models instead of real human-created data. During this recursive feedback loop, early stages of collapse result in the loss of low-probability edge cases and rare data distributions. As the training continues into late-stage collapse, the model's outputs homogenize, leading to repetitive, bland responses, increased hallucinations, and eventual structural degradation of the model's output variance.
Why this matters for your business
As AI-generated content increasingly floods the internet, preventing model collapse is vital for AI developers. It is crucial for ensuring the long-term viability and accuracy of future foundation models.
