What is synthetic data?
Training data generated by a model rather than written by a person. Human reviewers check and correct it before it is used.
Synthetic data is training data written by an AI model instead of by a person. A lab might ask a strong model to generate thousands of maths problems with worked solutions, customer support dialogues, or code with tests, then use the output to train another model.
Why it needs people
Model-written data repeats the generating model's mistakes and blind spots. Left unchecked, a model trained on it can learn wrong facts with confidence. So most synthetic data passes through human review before it is used: people check correctness, remove duplicates and near-duplicates, and fix or discard bad examples.
What the work looks like
Review tasks show a generated item and ask whether it is correct, realistic and useful, often with a corrected version required when it is not. Experts are hired for domains where errors are hard to see, such as advanced maths, chemistry or law. Some projects also ask you to write seed examples that the model then imitates at scale.
Example
A model generates 5,000 physics word problems. A reviewer finds that one in eight has a solution that silently drops a unit conversion. The faulty items are corrected, and the pattern is reported so the generation prompt can be fixed.
More terms: how models are trained
Full glossaryReady to look at roles?
Every listing shows its pay and who it is open to.