Synthetic data
- Training
- Standard term
Synthetic data is data made by a program, such as an AI model or a simulator, instead of collected from events as they happened. It is often used for training, testing, filling gaps, or privacy.
The “generator,” the program making the data, produces examples on request, so a team can get thousands quickly and ask for the rare cases its records barely hold. The examples also carry the generator's habits, including its favorite phrasings, its mistakes, and its blind spots.
Where you’ll meet it
You're setting up an appointment chatbot for a dentist's office. You have a few dozen past requests, so you ask an AI to write several hundred more, covering cancellations, insurance questions, and families booking together.
What to do
Use the generated requests to find where the bot stumbles. Before you trust it with patients, test it on a separate set of requests patients wrote, with their names removed, and keep that set out of everything you use to tune the bot.
Easy to mix up
Anonymized data keeps each original record, minus names and identifying details. Synthetic data keeps only the patterns, though now and then a generated record copies an original too closely.
Check yourself
You ask an AI for 1,000 sample customer complaints to test a tool that sorts complaints by topic. Compared with complaints customers send, how will the batch tend to read?
Go deeperMuch of a model's later training now runs on generated data
In post-training, the stage after a model's first training run, models write practice problems, answers, and conversations; other models or automated checks filter them; and what survives becomes training examples. A strong model's outputs can also train a smaller one, a method called “distillation.”
Generators built to protect privacy use methods such as differential privacy, which adds measured noise during training to limit how much any one person's record can affect what comes out. The noise sets a tradeoff, since stronger protection makes the generated data less faithful to the original records.