Synthetic data is artificially generated data that preserves statistical properties of a real dataset without containing real records.
Where it helps
Development without personal data. A team can build and test a system on synthetic data, without developers having access to real people's data. That is a genuine risk reduction.
Rare cases. If an event appears ten times in the entire history, the model has nothing to learn from. Generating plausible variations can help — carefully.
Testing the edges. You can deliberately construct extreme cases or combinations that do not occur in the data, to see how the system behaves.
Sharing data with third parties. A partner or supplier can work on a synthetic set without any transfer of personal data.
Where it misleads
Evaluating on synthetic data. If the generator was built on the same assumptions as the model, the evaluation measures the coherence of those assumptions, not real performance. The results are good and mean nothing.
Rare cases, again. Synthetic data can reproduce what has been seen before. It cannot invent a type of event nobody has ever observed — precisely the one you would want it to cover.
The illusion of anonymisation. A synthetic set generated from a very small source, or reproducing structure too faithfully, can allow inferences about real people. \u201cSynthetic\u201d does not automatically mean \u201crisk-free\u201d.
Amplifying bias. The generator learns from real data, imbalances included. A larger synthetic set can make an imbalance look like a solid regularity.
The practical rules
- Always evaluate on real data too, even on a small sample. Synthetic data can be most of training, but it cannot be the only test.
- Document how it was generated. A synthetic set without a methodology is impossible to interpret.
- Check the re-identification risk before treating it as non-personal data, particularly when the source set is small.
- Do not report results from synthetic data as system results without saying so explicitly.
Synthetic data is a useful tool. It becomes a problem the moment it is used to avoid the very check it should have made easier.