Skip to content

Glossary

Synthetic data

Synthetic data is training material generated rather than collected, used where real examples are too scarce, too sensitive or too expensive to label in the quantity a project needs.

In plain terms

Made-up examples, produced on purpose. If you need a thousand pictures of a rare fault and have eleven, or a thousand customer records you are not allowed to use, somebody generates material resembling the real thing and trains on that. It works better than it sounds and it inherits every blind spot of whatever produced it.

01

Why it matters

The scarce-example problem is a frequent reason a promising project stalls, and this is the standard answer to it. Rare faults, unusual claims and edge cases are by definition the ones you have least of and the ones you most need a system to handle. It also gives a route around material you cannot lawfully use for training, which for regulated organisations is often the difference between a project existing and not.

02

How it works

Material is produced by a model, by a simulation, or by transforming real examples into plausible variations, and which of those you use depends on the medium. Images tend to come from transformations and simulations; text more often from a model asked to produce examples in the style of the real ones.

It works best filling gaps rather than replacing the collection entirely. A modest quantity of real examples anchors the exercise, and generated variations extend the coverage into cases you have too few of. Systems trained wholly on generated material are the ones most likely to be excellent in tests and disappointing in use.

It inherits the limitations of whatever produced it, which is the central caution. Generated from a model, it carries that model's gaps and habits. Generated from a simulation, it carries every simplification the simulation makes. The system then learns those alongside the intended lesson and nobody notices until real material arrives.

The privacy claim needs care rather than acceptance. Material generated from real records can retain enough of them to be re-identifiable, so synthetic and anonymous are not the same word. Where the point of the exercise was avoiding an obligation, that has to be established properly rather than assumed from the label.

The two things it is asked to do

The two things it is asked to doBoth columns describe generating material and only one of them produces a system that works. On the left a small real collection sets the target, generated variations extend it into the cases you have too few of, and the result is measured against real examples deliberately held back from training. That last step is what makes the whole exercise honest, and it is the step most easily skipped. On the right nothing real is involved at any point, so the system learns whatever the generator believed the world looked like, and the test confirms it learnt that successfully. The results are excellent and mean nothing, and the discovery happens when real material arrives. Anchoring in something real, and testing against something the generator never saw, are the two habits that separate the columns.Filling a gapEleven real examples of a rarefault.Generated variations extend thecoverage.Tested against held-back realones.Usually works.Replacing collectionNo real examples gathered.Everything produced by agenerator.Tested against more generatedmaterial.Excellent in testing, poor inuse.The same technique, used twoways, with opposite outcomes.The difference is not thegeneration but whether anythingreal anchors it, and whether thetest used something thegenerator never saw.
Both columns describe generating material and only one of them produces a system that works. On the left a small real collection sets the target, generated variations extend it into the cases you have too few of, and the result is measured against real examples deliberately held back from training. That last step is what makes the whole exercise honest, and it is the step most easily skipped. On the right nothing real is involved at any point, so the system learns whatever the generator believed the world looked like, and the test confirms it learnt that successfully. The results are excellent and mean nothing, and the discovery happens when real material arrives. Anchoring in something real, and testing against something the generator never saw, are the two habits that separate the columns.
03

Seen in the wild

  • Ask an assistant to produce fifty realistic examples of a support enquiry you rarely receive, which is the text version of this at its simplest.

    ChatGPT
  • Generate variations of an awkward document to test an extraction step, where the point is coverage of the difficult shapes rather than realism.

    Claude
  • Build a step that produces test cases automatically, so a system gets exercised on the rare situations rather than only the common ones.

    n8n
04

Common misconceptions

People assume

It is automatically private.

In fact

Material generated from real records can carry enough of them to be traced back, so synthetic and anonymous are different claims. Where avoiding an obligation was the purpose, that needs establishing properly, with advice where the obligation is a legal one, rather than being inferred from the word.

People assume

It can replace collecting real examples.

In fact

It extends a real collection far better than it substitutes for one. A system trained wholly on generated material learns whatever produced that material, including its blind spots, and the results are frequently excellent in testing and poor in use. Real examples anchor the exercise, even a modest number of them.

05

Telling them apart

Synthetic data vs Training data

Synthetic data

Generated deliberately for a purpose, in the quantity you need, with the gaps of whatever produced it.

Training data

Collected from the world, uneven, and carrying whatever the world happened to contain.

One has the gaps of its generator and the other has the gaps of reality. Neither is neutral, and knowing which kind you are holding tells you which gaps to look for.

06

Questions

When is it the right answer?
When real examples are genuinely scarce, when the cases you need most are the rarest, or when the material you have cannot lawfully be used for training. Those three cover most legitimate uses. Generating material because collecting it is tedious is the version that disappoints.
How do we know it is any good?
Hold back real examples, train on the generated material, and test against the real ones you kept. That is the only measurement that means anything, and it is worth doing before rather than after the project depends on the answer. Testing generated material against generated material tells you nothing.
Does it satisfy our privacy obligations?
Take that to whoever advises you on the obligation, and bring two specifics: what the generator was trained on, and whether any single real record can be matched to an output. Those are the two things such an assessment turns on, and neither is settled by the word synthetic appearing in a product description.
07

Key takeaways

  • It is generated training material, used where real examples are scarce, sensitive or too costly to label.
  • It works best extending a real collection rather than replacing one.
  • It inherits the gaps of whatever produced it, and the system learns those alongside the intended lesson.
  • Synthetic and anonymous are different claims, and the second needs establishing.
  • The only meaningful test is against held-back real examples.
09

Tools that use this

  • ChatGPT

    Generating examples of a rare enquiry, the simplest version of this.

  • Claude

    Variations of an awkward document, for coverage rather than realism.

  • n8n

    Producing test cases automatically, so the rare situations get exercised.

Last checked July 2026

All glossary terms