What Synthetic Data Is For

Synthetic data fills gaps that real data leaves behind. It inherits the standards of whoever designed it, which is both its strength and its limit.

5 August 20260:10Data, Model Development, Artificial Intelligence
00:00
0:10
Nicole Junkermann beside jars of coloured material while recording an AI Overview podcast briefing on synthetic data

Synthetic data means carefully made examples used to train or test systems. It can fill gaps, but quality matters. Artificial examples still need human standards.

Full transcript of Briefing 06, 0:10, published 5 August 2026.

  • Synthetic data is constructed, not collected
  • Its main use is covering cases real data under-represents
  • It inherits the assumptions of whoever specified it
  • Generated volume is not the same as generated quality

Real datasets are uneven. The common cases are heavily represented and the rare ones — the edge conditions that matter most when a system fails — are often nearly absent. Synthetic data is a way of deliberately manufacturing those under-represented cases so a system can be trained or tested against them.

Used this way it is a targeted instrument, not a substitute for collection. The question it answers is narrow: what situations does this system need to have seen, that the real record does not contain enough of?

Because synthetic examples are designed, they carry the designer's assumptions about what a realistic case looks like. If those assumptions are wrong, the result is a system confidently trained on a world that does not exist — and the error is harder to spot than a gap would have been.

That is why volume is a poor measure of success here. The useful checks are whether the generated cases were specified by someone who understands the domain, and whether the system is still evaluated against genuine data before it is relied upon.

What is synthetic data?

Purpose-built examples used to train or test a system, rather than examples collected from real events. It is constructed to specification instead of observed.

Why use it at all?

Chiefly to cover situations that real datasets under-represent, including rare failure conditions, and to work on a problem where real records are restricted or sensitive.

What is the main risk?

That the generated examples reflect an inaccurate idea of the real world. A system trained on them can look well-tested while remaining unprepared for the cases that actually occur.