Selling to Labs
Synthetic didn't replace you
Every data owner hears it, usually from someone with no stake in the answer: the labs can generate their own data now, so yours is worth less. It is a reasonable-sounding claim built on a category error. Synthetic data is a transformation applied to something a model already knows or can verify. It expands, rephrases, and searches — it does not observe. The question is not whether synthetic works. It is what it works on.
What synthetic data actually is
| Kind | What it does | What it needs to already have |
|---|---|---|
| Rephrasing / augmentation | Restates existing documents in cleaner or more instructional form | The source documents |
| Distillation | Transfers a stronger model's behaviour into a weaker one | A stronger model that already has the capability |
| Self-play with a verifier | Generates attempts and keeps the ones a checker accepts | A domain where correctness is programmatically checkable |
| Free generation | Produces plausible text in a domain with no grounding | Nothing — and this is the kind that degrades models |
Three of those four require real data or a real verifier as an input. Only the fourth is the thing sellers are told will replace them, and it is the one that does not work.
What the collapse literature supports
The findings are more nuanced than either camp quotes them as being, and both numbers matter to a seller:
- Degradation becomes severe somewhere around a 30–50% synthetic share for several studied backbones when synthetic data is fed back recursively. Distribution tails thin out first; the model becomes fluent, generic and progressively worse at rare cases.
- Measurable degradation appears far earlier than that — contamination on the order of 1–10% has been enough to produce measurable harm in code and vision settings. The effect is not a cliff at 50%; it is a gradient that starts low.
- Grounding prevents it. Where synthetic data is anchored — to real documents, to a verifier, to a stronger model with genuine capability — the collapse dynamic largely does not appear. Accumulating real data alongside synthetic, rather than replacing it, is what breaks the feedback loop.
Read together: synthetic data is safe in proportion to how tightly it is anchored to something real, and dangerous in proportion to how free it is. Which makes real data the input that governs how much synthetic can be used at all.
The 2026 pattern
What the fully-open recipes actually do — and what the closed labs' published work implies they do — is layer synthetic on top of licensed and curated real data, at every stage:
- Real corpus in. Filtered, deduplicated, quality-classified.
- Synthetic rephrasing of that corpus into denser, more instructional form. The facts come from the real documents; the model supplies the phrasing.
- Synthetic expansion around real seed examples — variations on real cases, not invented ones.
- Verifier-filtered generation for domains where a checker exists, keeping only what passes.
- Real held-out data for evaluation, because a model evaluated on its own output measures nothing.
At every one of those steps, more and better real data raises the ceiling of what synthetic can produce. This is why demand for licensed data went up over the same period synthetic techniques matured. The multiplier got better, and a better multiplier makes the multiplicand more valuable, not less.
Where synthetic genuinely does replace you
Being honest about this is what makes the rest credible. If your corpus is any of the following, synthetic generation is a real substitute and your price should reflect it:
- Format conversion at scale — turning documents into question-answer pairs, summaries, or instruction pairs. A model does this now, cheaply, and buying it is hard to justify.
- Common-knowledge instruction data — general assistant behaviour, standard tasks in well-covered domains.
- Problems with mechanical answers in domains models already handle — routine arithmetic, standard code transformations, template-driven text.
- Paraphrase volume — more ways of saying what is already in the corpus.
What is not substitutable: observation of the world (what actually happened, to whom, with what outcome), expert judgement on genuinely contested cases, the real distribution of requests a professional receives, failures and recoveries, and anything whose correctness depends on facts the model cannot access.
How to answer the objection in a meeting
Not by arguing about collapse papers. By locating your corpus on that split:
"Everything in this corpus is an observation with an outcome attached. A model can generate a plausible triage conversation. It cannot generate what the patient actually turned out to have — and that field is why the corpus exists. Here is what happens to the eval when you drop it."
Then show the measurement. A contribution test against a matched control settles this argument in a way no amount of literature citation will; that is what marginal is for, and if your data really is substitutable, the test will say so — which is worth knowing before a buyer's pilot says it for you.
A related consequence: if you have used synthetic augmentation on your own corpus, disclose the share and the method. Buyers screen for it, undisclosed synthetic content is treated as a hygiene failure, and the detection question is its own post.