ARTIFICIAL INTELLIGENCE
The internet’s high-quality data is effectively exhausted for training frontier AI models. That’s the consensus. The proposed solution is synthetic data — and it deserves more scrutiny than it’s getting.
The idea is elegant: if you’ve run out of real data to train on, use AI to generate synthetic data that captures the same statistical properties. Train the next generation of models on a combination of real and synthetic data. Problem solved.
Except it’s not quite that simple.
What Synthetic Data Is Good At
Synthetic data genuinely works well for specific, structured tasks where the distribution of correct answers is well-defined. Mathematical proofs, coding problems, logic puzzles, scientific calculations — these can be synthetically generated at scale and used to meaningfully improve model performance on those tasks. The reasoning model improvements of 2025 were largely built on synthetic data of this type.
“Training a model on its own outputs is like photocopying a photocopy — the errors compound while the signal degrades.”
The Model Collapse Problem
The deeper issue is model collapse — what happens when you train models on synthetic data generated by earlier models. Training a model on its own outputs is like photocopying a photocopy. The errors compound, the edge cases get smoothed away, the diversity of the distribution degrades, and you end up with a model that is less capable than what you started with in ways that are subtle and hard to detect.
Research on this is still active, and the magnitude of the problem is contested. But the risk is real, and the field’s confidence that synthetic data simply solves the training data problem deserves more scepticism than it’s currently receiving.
Tags: Artificial Intelligence • Opinion • Technology & Society • 192.168.1.22/