Key points
- Synthetic data is data that didn't happen for real — it got made up because it was needed.
- There are three main reasons to make it: the real thing is scarce, it's personal, or labeling it is costly.
- It can be produced in any quantity, and when built from rules or a simulation the answer comes along with it, so far less time gets spent labeling it.
- In exchange, it drifts from the real thing in spots. Whatever the rules or source data didn't contain rarely shows up in it.
- Making more of it by copying earlier made-up batches drifts further from reality each round, which is why real data still gets mixed back in.
The analogy
There are two ways to lay down a field. One is planting real grass. It takes time to root, gets torn up by rain, and wears down under heavy use. All that upkeep buys a feel that nothing else quite matches.
The other is laying artificial turf. Order however much is needed and it goes down in a day, holds up in the rain, and shrugs off heavy use. Synthetic data is that turf.
Turf drifts from the real thing in spots too. A ball rolls at a different speed, and a slide feels different underfoot. A team that only ever practiced on turf finds its footing off the moment it steps onto real grass. And turf's texture was designed by copying real grass in the first place — with no real grass to copy, there'd be nothing to imitate at all.
In detail
Why make it up instead
The most common reason is scarcity. The rarer something is, the harder it is to collect. The exact moment a machine fails, or a genuinely dangerous situation, doesn't happen often — yet a model that handles it well is exactly what's needed.
The second reason is that the data carries personal information. In fields where real records can't be used as they are, data built to match only the statistical shape of the real thing sometimes gets used instead. Making sure no individual record can be traced back is the whole point here.
The third reason is the cost of labeling. Having a person tag the correct answer by hand takes time and money. Data made from rules or a simulation already comes with the answer known, since whoever made it knows exactly what they made. Data made by a model can carry wrong answers, though, so a separate filtering step is kept.
How it gets produced
The simplest approach writes down rules — what range a value falls in, how it behaves — and draws straight from those rules. It's easy to work with because exactly what went into it is known, but any property the rules didn't mention simply never shows up in the data.
Another approach learns from real data and produces something similar to it. Images and text lean on this approach heavily. It gets closer to the real thing, but whatever quality the source data had carries straight through into the copy, for better or worse.
This is worth separating from lightly altering data that already exists. Flipping a photo or adjusting its brightness keeps the original intact underneath; synthetic data is the other direction — making something that wasn't there before.
Knowing where it drifts is the real skill
Made-up data only holds the world its maker imagined. Exceptions the rules never covered, the odd rare values that show up in real data, the strange things people actually type — most of that is missing. A model trained purely on data like this tends to do well in a tidy setting and stumble the moment it hits the real world.
Sameness is a trap too. Turf feels the same wherever it's stepped on; real grass varies patch by patch. Made-up data often ends up too similar to itself — the count grows while what there is to learn from it doesn't. That's why how different the examples actually are matters more than how many there are.
Final grading, for that reason, happens on real data. Training and grading both on made-up data can look like success, but that score only holds inside the maker's own imagination.
What happens when a copy gets copied again
The thing to watch most closely is feedback. A model trained on made-up data goes on to make more data, and if that cycle repeats, the result drifts from reality a little more each time — a photo of turf feeding the design of the next batch of turf, feeding the one after that.
The drift follows a pattern too. The rare cases disappear first — only the common shapes are left once the corners get trimmed away. What remains then grows more similar to itself over more rounds. That's why real data gets mixed back in at a steady share whenever this is used for real, with a record kept of how much of the total is made up.
More precisely
Synthetic data doesn't erase privacy concerns entirely. If the model producing it memorizes a specific real record too closely, something very close to that record can surface again in what it generates. Because of that, checking how much overlap exists between the made-up data and the original is a separate step that has to happen.
The analogy has a limit too. Artificial turf is visible to the eye as artificial; well-made synthetic data isn't easy for a person to tell apart from the real thing. Since the drift isn't visible, it gets checked indirectly instead — plotting the made-up data's distribution alongside the real one, or grading on real data alone.
Where it earns its keep is worth knowing too. Synthetic data is most trustworthy when it is built on top of real data. These days a large share of language-model training data is sometimes filled with synthetic data, but even then real data serves as the model to copy and a filtering step is kept. Laying it down wall to wall where no real data exists at all removes any way to check what it's imitating, which means there's no way to measure how far it has drifted either. That's a large part of why noting how much made-up data went in, and what it was modeled on, is a documentation practice that gets recommended.
Try it yourself
Common misconceptions
It's easy to think more synthetic data is always better, but actually when the examples are too similar to each other, the count grows while what the model actually learns doesn't.
It's easy to think made-up data is free of bias, but actually the leanings of the source data and the rules its maker chose carry straight through into it.
It's easy to think using synthetic data erases privacy concerns, but actually a memorized fragment of the original can surface again, so the overlap has to be checked separately.
One-line summary
Synthetic data is data made up because it was needed, good for filling a scarce spot, but final grading has to happen on the real thing.
Spotted an error or have a better analogy? Suggest an edit · Last updated 2026-10-18