Model collapse is a degenerative phenomenon in machine learning where AI models trained on data generated by other AI models progressively lose quality, diversity, and accuracy over successive generations. The process creates a feedback loop in which each generation of model amplifies the errors and biases of the previous one, ultimately producing homogenized, error-prone outputs disconnected from real-world data distributions. ¹⁾

The phenomenon was formally identified and named in the landmark 2023 paper “The Curse of Recursion: Training on Generated Data Makes Models Forget” by Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, and colleagues, initially published as a preprint and later in Nature (2024). ²⁾ The researchers demonstrated that training generative models — including language models, variational autoencoders (VAEs), and Gaussian mixture models — on their own outputs causes compounding information loss across generations.

The paper identified two distinct stages:

• Early model collapse — the model loses information from the tails of the data distribution, such as minority classes and rare events. This stage is often unnoticed because overall performance metrics may appear stable or even improve while rare but important data points are forgotten. ³⁾
• Late model collapse — severe performance degradation occurs, with the model producing confused concepts, dramatically reduced variance, and outputs that drift into repetitive or nonsensical content. ⁴⁾

The core mechanism of model collapse is a recursive feedback loop. When AI-generated content is published to the internet and subsequently scraped for training data, new models are inadvertently trained on synthetic rather than human-generated data. Each successive generation:

1. Under-samples low-probability events — rare but important data points are progressively excluded
2. Over-samples common patterns — majority distributions are amplified with each generation
3. Accumulates errors — small distortions compound like a game of telephone, where each retelling introduces additional noise ⁵⁾

The analogy commonly used is photocopying: each copy of a copy degrades the image until details are lost entirely. ⁶⁾ In a concrete example, a dataset with 90% yellow objects and 10% blue objects produces a model that generates an even higher proportion of yellow. After several generations, blue objects disappear entirely from the output. ⁷⁾

Researchers have termed this process model autophagy disorder (MAD), or colloquially, AI cannibalism — a state where models consuming their own outputs produce increasingly homogenized results detached from reality. ⁸⁾

Model collapse makes data provenance — tracking the origins and history of training data — a critical concern. As AI-generated content floods the internet, distinguishing synthetic data from human-created data becomes increasingly difficult. ⁹⁾

Key challenges include:

• Contamination at scale — with generative AI producing vast quantities of text, images, and other media, web-scraped training datasets are increasingly contaminated with synthetic content
• Bias amplification — uncurated synthetic data perpetuates and magnifies existing biases and stereotypes present in the original training data ¹⁰⁾
• Ground truth erosion — the fundamental reference points against which model accuracy is measured become corrupted when they are themselves AI-generated

The Harvard Journal of Law and Technology has noted the legal implications, arguing for a “right to uncontaminated human-generated data” as a foundation for maintaining AI quality. ¹¹⁾

Researchers have proposed several approaches to counter model collapse:

• Human data preservation — maintaining curated reservoirs of verified human-generated data for training
• Provenance tracking — implementing systems to tag and filter synthetic content from training pipelines
• Data diversity enforcement — actively ensuring training datasets represent the full distribution of real-world phenomena, including rare events
• Hybrid training — combining synthetic and human data with explicit controls on the proportion of each

Model collapse intersects with several related concepts in machine learning:

• Catastrophic forgetting — the tendency of neural networks to lose previously learned information when trained on new data
• Data drift — changes in data distributions over time that degrade model performance
• Overfitting — models that perform well on training data but fail to generalize

• AI Slop
• Digital Pollution
• Homogenization of Expression