Model collapse is a degradation process that occurs when an AI model is trained predominantly on data generated by other AI models rather than on original human-produced content. Each generation of training amplifies the statistical biases and omissions of the previous model's outputs, causing the distribution of the learned data to narrow: rare but legitimate content disappears, and the model's outputs become increasingly homogeneous or distorted. Research published in 2023 demonstrated that even a small fraction of AI-generated text in training data, if compounded across multiple generations, causes measurable degradation in language diversity and factual coverage.

Think of it like successive photocopies of a photocopy: each copy introduces small distortions, and after enough generations the image loses fine detail until only a blurry approximation of the original remains.

Controlled experiments have shown model collapse in language models trained iteratively on their own outputs. Quality metrics drop predictably:

Generation 0 (human data):   perplexity = 18.3, vocab diversity = 0.91
Generation 3 (AI-on-AI):     perplexity = 24.1, vocab diversity = 0.73
Generation 8 (AI-on-AI):     perplexity = 41.7, vocab diversity = 0.44

Mitigation strategies include maintaining a fixed proportion of verified human data in every training run, using data provenance tracking to flag synthetic content, and applying deduplication pipelines before training.

Model collapse connects directly to Pretraining data quality, Overfitting (a related failure of generalisation), Hallucination (which worsens as diversity collapses), and the emerging field of synthetic data management in LLM pipelines.

You will encounter model collapse discussions most often in debates about the long-term sustainability of web-scraped training data as AI-generated content saturates the public internet.