This is the first part of a two-part series. The second, “Evolution with Memory,” examines the mathematics of collapse.
An alarmist story has been gaining ground in the tech industry: artificial intelligence models are said to be on the verge of an irreversible degeneration known as model collapse. According to this hypothesis, rigorously formulated by Shumailov and colleagues in Nature (2024), a system trained recursively on content produced by other systems progressively loses its nuance and falls into a loop of decay, much like making a photocopy of a photocopy.
Presenting this scenario as an unavoidable apocalypse, however, is a simplistic and hasty reading.
A Hypothesis Under the Magnifying Glass
If the collapse theory were infallible, the software industry would be stuck in a dead end. But the argument assumes laboratory conditions that rarely occur in the real world.
For a model to collapse that drastically, it would need total isolation from human data and a complete absence of oversight in filtering information. Does anyone believe that the big tech companies would let their models swallow every scrap of text on the web without any prior filtering?
It is hardly plausible that labs like OpenAI, Google or Anthropic would train their next models by feeding them uncontrolled digital garbage. And there is something else: the human corpus that predates the explosion of generative models does not disappear. It is archived, and nobody throws it away when a synthetic text appears.
The Value of High-Quality Synthetic Data
The premise that all AI-generated data is “toxic” for future generations of models ignores the role of curated synthetic data. In fields such as mathematics, programming or chess, data generated by models under strict logical rules has proven useful, and sometimes better than much human data riddled with errors and biases.
If AI could only degenerate by learning from itself, there would be no way to explain how AlphaZero reached a superhuman level by playing millions of games exclusively against itself. The key lies not in where the data comes from, but in the ability to verify its quality and coherence.
Still, the example should not be stretched too far. Chess has a perfect verifier: the rules and the outcome of the game. Natural language has none, and that is where the hard part of the problem lies.
An Engineering Problem or an Insurmountable Limit?
Rather than a systemic catastrophe, model collapse is a data engineering challenge. Curation, digital watermarking and synthetic-content classifiers are tools for deciding which information to keep and which to discard, although today they are far from infallible, especially for text.
Believing that artificial intelligence will “self-destruct” for lack of original human text is an overblown interpretation. The technology is not doomed to collapse. What it is bound to change is the way the industry selects, validates and synthesizes the information with which it builds the future.