Richard Sutton, the reinforcement-learning pioneer who shared the 2024 ACM Turing Award with Andrew Barto, has called the AI industry's growing reliance on synthetic data "a big mistake."
According to The Decoder, Sutton's objection isn't that synthetic data is worthless — it's that treating it as a scalable stand-in for real-world experience misreads the size of the problem AI is trying to solve. He frames the issue against what he calls an infinitely complex world; in our reading, the underlying argument is that no dataset generated by a model can meaningfully extend past the boundaries of what that model already knows.
The timing matters. As labs run low on high-quality human-generated text to train on, synthetic data — text, code, and dialogue produced by existing models to train the next generation of models — has become one of the industry's default answers to the data-scarcity problem. Sutton's pushback lands as a direct challenge to that default.
A pioneer whose own ideas fueled the scaling era
Sutton's credibility on this question comes from an unusual angle. He co-wrote the field's standard textbook on reinforcement learning with Barto, and his 2019 essay "The Bitter Lesson" became one of the most cited arguments for the current era of AI: general methods that exploit ever more computation and data, the essay argued, beat approaches built on hand-crafted human knowledge. That essay is routinely invoked to justify scaling up training data by any means available — including synthetic generation.
Sutton turning that same scaling logic against synthetic data specifically is notable, not because he is reversing course on scale, but because he's drawing a line between more data and more of the right kind of data. In our reading, his argument treats model-generated data as a closed loop: it can only reflect and recombine what the generating model already encodes, rather than introduce genuinely new signal from the world.
Why "infinitely complex" is the crux
The phrase Sutton uses — an infinitely complex world — points to a structural limit rather than a quality-control problem. Synthetic datasets, however large, are bounded by the distribution of the model that produced them. Real environments keep generating edge cases, rare events, and novel combinations that no generator trained on past data will anticipate. This is consistent with a body of research on "model collapse," where models trained repeatedly on their own or on other models' outputs tend to narrow toward the statistical center of their training distribution, losing coverage of the tails over successive generations.
Sutton's own research agenda — built around agents that learn continually from direct interaction with an environment, rather than from a static, pre-packaged corpus — offers a clue to what he'd prefer instead: grounding in real, ongoing experience over any amount of pre-generated text.
What this means for teams building on synthetic pipelines
None of this makes synthetic data useless for practical engineering — it remains a cheap, controllable way to cover known gaps. But Sutton's critique is a useful checklist for anyone leaning on it as a primary data strategy:
- Track the synthetic-to-real ratio in each training run, not just the total token count — a growing share of synthetic data can quietly narrow what the model actually knows.
- Watch for distribution collapse across successive fine-tuning rounds, especially when synthetic outputs from one model version feed the next.
- Keep a real-world evaluation set that synthetic data was never allowed to touch, so degradation on genuine edge cases doesn't hide behind strong benchmark scores.
- Treat synthetic generation as augmentation, not replacement, for domains where the underlying environment keeps producing genuinely new situations — customer interactions, live agentic tasks, real sensor data.
AiiN's takeaway
Sutton's comments are a useful counterweight at a moment when synthetic data has become close to consensus practice across the industry. Nobody credible is proposing that teams abandon synthetic data outright — it's too cheap and too useful for that. But if you're designing a pipeline that leans on it heavily, Sutton's warning is worth keeping in view: a model can only generate variations on what it already knows, and a training strategy built entirely on that loop will eventually run into the same wall he's describing — a world more complex than anything the loop can produce on its own.