Author(s): Rohan Mistry

Originally published on Towards AI.

AI is training on AI. Nobody’s talking about it loudly enough.

Something is quietly breaking inside the world’s most powerful AI systems.

After posing the core question of what AI is “learning” as the internet fills with AI-generated content, the article argues that training can turn into a recursive loop: models learn from increasingly synthetic outputs, which flattens rare knowledge and erodes authenticity. It explains model collapse as a probabilistic process where edge cases get suppressed over successive generations, producing fluent but incomplete systems that “don’t know what they don’t know.” The damage may become permanent because contaminated information is baked into model weights as persistent “poisoned priors,” making it hard to patch without retraining on fresh, sufficiently large, uncontaminated human data. The author also weighs counterarguments that careful mixing and multimodal training might mitigate collapse, but contends that current incentives favor cheap synthetic content and weaken provenance and curation. Through examples, including a scenario where AI-assisted literature leads to missing rare medical knowledge, the piece highlights how systems can fail invisibly. It concludes by reframing the problem as one of “data wars” and scarcity: labs increasingly pay for provenance and authentic human writing, while synthetic data can be an excellent supplement but becomes dangerous as a replacement, since human expression carries contradictions, context, and lived texture that synthetic training alone cannot reproduce.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI