The training data behind today's best models looks very different from what powered the breakthroughs of 2022 and 2023. Most of it is now synthetic — generated by models, filtered by models, and curated by models — and that shift is one of the most important and least-discussed developments in frontier AI.
Why the shift happened
The supply of high-quality human-written text on the open web has not grown as fast as the demand for training data. The easiest path to more data — just scrape more of the internet — produces rapidly diminishing returns. Each new model needs more and better data than the last, and the open web simply does not contain enough of it.
Synthetic data solves this in two ways. First, it can be generated in essentially unlimited quantities. Second, and more importantly, it can be generated to target specific weaknesses — hard reasoning, rare languages, domain expertise — that the open web underrepresents.
How it works in practice
The dominant pattern is to use one model to generate training data for another. A frontier model might generate millions of examples of a specific kind of reasoning, which are then filtered for quality — often by yet another model — and used to fine-tune a target model on that capability. The pattern is recursive: better models produce better synthetic data, which trains the next generation of better models.
Verification is the key technical challenge. Synthetic data is only useful if you can be confident it is actually correct. The leading labs have invested heavily in automated verification techniques: running generated code, checking math derivations symbolically, cross-referencing factual claims against trusted sources, and using ensembles of models to catch errors.
The reasoning revolution
The single biggest win from synthetic data has been in reasoning. The current generation of reasoning models was trained on enormous quantities of synthetic reasoning traces, generated by earlier models and filtered for correctness. This pattern is what produced the jump in math, code, and logic capability that defined the 2025-2026 model generation.
The risks
Model collapse — the phenomenon where models trained on their own output gradually degrade — was a serious worry a few years ago. The actual empirical picture has turned out to be more nuanced. With careful filtering, diverse generation, and continued grounding in human-produced data, the leading labs have largely avoided collapse so far. But the dynamics are not fully understood, and complacency would be premature.
A subtler risk is homogenization. If most synthetic data comes from a small number of similar source models, the resulting training data may inherit the biases, blind spots, and quirks of those models. The eventual generation of trained models could be more similar to each other than is healthy for the field.
The legal and ethical questions
Synthetic data partially sidesteps some of the legal questions around training data, since the synthetic text is technically generated by the model rather than copied from a human source. But the question of whether the source models that produced the synthetic data were themselves trained on legitimate sources is unresolved. The legal picture here will get more interesting before it gets simpler.
On the ethical side, the use of human contractors to generate the seed data and verify the synthetic outputs has become a significant industry. Working conditions, compensation, and the psychological impact of this work — especially the parts that involve filtering harmful content — are receiving overdue attention.
Where this is going
Synthetic data will only grow in importance. The trajectory points toward a world where the majority of training data for frontier models is generated, where human-written text plays the role of grounding and validation rather than bulk, and where the quality of a lab's synthetic data pipeline is one of its most important competitive advantages.
For everyone outside the frontier labs, this trend matters because it shapes the capabilities you will be able to access. The next generation of models will be smarter, more capable, and more diverse in their strengths than the current generation, largely because of synthetic data techniques that are still being invented as you read this.