Put a photograph through a photocopier. Copy the copy, then copy that copy. At first the page looks fine. A few generations later, pale details disappear, dark edges thicken, and a dust speck has somehow become a permanent family member.
Synthetic data can behave like that. It is cheap to make, easy to shape, and enormously useful. It is also generated from a model's existing picture of the world. Calling the output “new data” does not make it a new observation.
More copies, not more world
Suppose a classifier struggles with damaged packages. A generative model can create ten thousand pictures of dented boxes by lunchtime. The dataset is larger. But its dents come from patterns the generator already learned. If real warehouse damage includes a rare kind of puncture absent from that model, no amount of sampling will photograph it into existence.
Synthesis expands a representation. Cameras, sensors, transactions, experiments, and people add evidence from outside it.
The tails fade first
The model-collapse experiments published in Nature examined what happens when models repeatedly train on model-generated data. Over generations, the learned distribution can lose information about the original one.
Common patterns survive because the generator produces them often. Rare patterns are easier to miss, then easier to miss again. This is exactly backward for many production systems, where the rare edge cases are why the model needs improvement.
Where synthesis helps
None of this makes synthetic data fake or useless. It shines when the target is explicit:
- generating variations around known failure cases,
- balancing classes while keeping a real validation set,
- simulating dangerous or expensive scenarios,
- distilling behavior from a capable teacher, and
- creating privacy-preserving development data.
In those cases, the synthetic examples are instruments. We know what gap they are meant to probe. The trouble begins when volume itself becomes the goal and the generator grades its own homework.
Keep the provenance
Label where each example came from, which model and prompt created it, what real seed it used, and which filters accepted it. Evaluate on untouched real data. Compare performance on rare slices, not only the average. Keep collecting fresh observations after launch.
A photocopier is a fantastic office tool. You just would not point it out the window and ask what the weather is. Synthetic data deserves the same clarity about its job.