synthetic data
/ sin-THET-ik DAY-tuh /
Synthetic data is artificially generated information made to stand in for real data. Instead of collecting more records from the world, you manufacture plausible new ones — using a simulator, a set of statistical rules, or a generative model that has learned what real examples look like. Think of a flight simulator: it produces realistic flying experience without anyone leaving the ground.
It is generated several ways. A physics simulator can render millions of labeled images of a part from every angle. A generative model trained on real patient records can spit out fictional but statistically similar patients. Simple rules can fabricate transactions that follow known patterns. Crucially, synthetic data can come pre-labeled — the simulator knows exactly what it drew — which sidesteps the slow, costly human annotation step entirely.
Why it matters: synthetic data shines when real data is scarce, expensive, dangerous, or private. It lets you train on rare events you can't wait for (a self-driving car practicing crashes), share data without exposing real people, and balance a rare class. But there is no free lunch: synthetic data can only reflect the assumptions or the training data it was built from, so it may miss the messy surprises of reality — a gap called the sim-to-real gap. Models trained purely on synthetic data often need real data to fine-tune and validate, and you must never assume synthetic looks exactly like the real thing.
A robotics team trains a warehouse robot inside a video-game-like simulator, generating millions of perfectly labeled scenes — boxes at every angle, every lighting — that would take years to photograph in real life. The robot then practices for an hour on real shelves to close the gap between the clean simulation and the messy warehouse.
Millions of free, perfectly-labeled simulated scenes — then a reality check.
Synthetic data inherits the blind spots of whatever produced it. If your simulator or generator never imagined a certain situation, your synthetic data won't contain it — and a model trained only on it will be confidently unprepared for the real world.