model collapse

Model collapse is what happens when AI grows by feeding on its own output, gradually becoming shoddier over time.

Think of photocopying a document, then photocopying that copy, and so on. The first few copies look fine, but with each repetition, fine details from the original slowly fade and blur. AI training works similarly. AI learns from text and images written by humans and produces its own output, that output piles up online, and when the next generation of AI trains on that pile, it ends up repeatedly learning AI's own averaged-out, flattened patterns instead of the rich variety of human originals.

For this worry to become real, there needs to be a steady supply of genuine new writing and art made by actual humans. But if writers and creators stop producing because they aren't compensated, that supply itself dries up. In the article's own words, this is an "AI famine" — a drying-up of human-written material for AI to learn from. Model collapse refers to the quality-decline scenario that results from it.

The article writes that "a writer who won a lawsuit raised against an AI company the scenario of model collapse, where training data gets contaminated by AI output and quality falls apart."

One thing easy to misunderstand: model collapse doesn't mean AI answers are already broken right now. It refers to a long-term risk that could unfold as human-made original content keeps shrinking, so in this article too, it appears not as the outcome of the copyright lawsuit victory but as a warning about what could come after.