Degenerative AI

2026-07-09 · Source: AI Advances - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Intermediate, quick

Summary

Degenerative AI describes a phenomenon known as "Model Autophagy Disorder" (MAD) or "model collapse," where generative AI models degrade in output quality and diversity when trained on data produced by other AI models. This process, likened to "mad cow disease" and cellular "autophagy," causes models to narrow their output distribution over generations. An experiment generating human faces demonstrated this, showing synthetic faces losing distinct features and converging on a "haunted average" after just a few generations. Ilia Shumailov and colleagues coined "model collapse" in their 2023 paper "The Curse of Recursion," further exploring it in 2024. The core issue is that models fed their own synthetic output learn from a diminishing pool of diversity, leading to a collapse of the entire distribution onto a single, artifact-ridden peak.

Key takeaway

For AI Engineers developing generative models, you must recognize the critical risk of "model collapse" when incorporating AI-generated content into training datasets. To prevent your models from becoming "degenerative" and losing output diversity, carefully curate your data sources. Ensure your training data consists predominantly of real-world, human-created examples to maintain robust model performance and avoid the convergence onto a single, artifact-ridden output.

Key insights

Generative models fed their own output degrade, leading to "model collapse" and loss of diversity.

Principles

• Generative models learning from synthetic data become degenerative.
• Training on self-generated data causes output distribution collapse.
• Model Autophagy Disorder (MAD) reduces output diversity over generations.

Method

The process involves a generative model repeatedly consuming and learning from its own synthetic outputs, causing its internal representation and subsequent generations to narrow and converge.

In practice

• Avoid training generative models on AI-generated content.
• Monitor output diversity across model generations.
• Identify "model collapse" by observing converging outputs.

Topics

• Generative AI
• Model Collapse
• Synthetic Data
• Data Contamination
• AI Model Training
• Output Diversity

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• The Day Synthetic Data Turned Poisonous: Inside Model Collapse — Recursive training on synthetic data causes model collapse, eroding diversity and amplifying errors.
• AI Is Eating Itself to Death — and Nobody Is Stopping It — AI models trained on synthetic data experience "model collapse," progressively losing unique information and degrading.
• From Collapse to Improvement: Statistical Perspectives on the Evolutionary Dynamics of Iterative Training on Contaminated Sources — Model collapse in iterative training can be avoided with fresh data and appropriate sample sizes.
• AI is Eating Its Own Tail And Biting The Hand That Feeds It — AI training on its own outputs creates a feedback loop leading to model collapse and degraded quality.
• Why Model Collapse In LLMs Is Inevitable With Self-Learning - Hackaday — LLMs, as statistical models, degrade into "model collapse" without continuous external human-generated data.
• Kim Crawley: Recovering After Gen AI — GenAI's bubble will crash due to unsustainability and model collapse, necessitating preparation for a post-AI era.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Advances - Medium.