This is the question that makes the 2026 accusations hard to adjudicate, and it does not have a clean answer.
The technical overlap
Distillation: query a model, collect outputs, train on them.
Synthetic data generation: query a model, collect outputs, train on them.
There is no procedural difference. The same API calls, the same corpus, the same fine-tuning step. If you were handed a training set and asked which it was, you could not tell.
The distinctions people draw
Three are commonly proposed, and each is leakier than it first appears.
Intent. Distillation targets a specific teacher's capability; synthetic data fills a gap in coverage. This is the most honest distinction and the least enforceable, because intent is not a property of the data.
Breadth. Distillation attempts broad capability transfer; synthetic data addresses a narrow task. Better, but it is a spectrum, and nobody has proposed a principled threshold.
Targeting. Distillation constructs prompts to probe the teacher's capability boundary; synthetic data generation simply generates useful examples. This one is the most operationally real, and it does leave a trace: systematic boundary-probing looks different from ordinary generation in traffic analysis. It is a large part of what the 2026 disclosures were describing.
Why the label does not change the legal position
The distinction matters less than people hope, because the applicable restriction is contractual and is written in terms of purpose, not vocabulary.
A prohibition on using outputs to develop a competing model attaches to what you do with the data. Relabelling a distillation corpus as synthetic data does not change whether you used it to train a competitor. Conversely, genuinely narrow synthetic data for a non-competing product is generally fine regardless of what you call it.
The contamination problem
There is a further complication that undermines the whole framing: a large and growing share of ordinary web text is model-generated. Any model trained on a recent crawl is training on other models' outputs whether or not anyone chose to.
This means unintentional second-order distillation is now the default condition of the field rather than an edge case, and it is a major reason behavioural fingerprints are unreliable. A model can absorb another system's characteristics without anyone having distilled anything.
Practical guidance
- Assume the restriction follows the data, whatever you name it
- Record where generated data came from. Provenance is what you will be asked for
- Narrow task-specific generation sits on much safer ground than broad capability probing
- If you would be uncomfortable explaining your prompt distribution to the provider, that is the signal, not the label you have chosen
Common questions
Is synthetic data the same as distillation?
Technically they overlap almost completely. Both train a model on text generated by another model. The distinction usually drawn is one of intent: distillation aims to reproduce a specific teacher's capability, while synthetic data generation aims to create training examples. Intent is not visible in the data itself.
Is it legal to train on synthetic data from a commercial model?
Copyright generally does not reach it, since model outputs are not copyrighted. The binding constraint is contractual. Most commercial terms prohibit using outputs to develop competing models, and that restriction attaches to the data regardless of whether you call it synthetic data or distillation.