In plain words
AI needs a lot of examples, and real ones are often a problem. They may be too few, like rare kinds of fraud. They may be private, like medical records. Or they may be dangerous to collect, like car crashes.
Synthetic data is the workaround. A program studies the real data, learns its patterns, and then produces new records, images or text that follow the same patterns but describe no real person or event. Simulators do the same for things like self-driving cars, generating endless virtual roads.
Today large language models are also used to write training data for other models: practice questions, worked answers, example conversations. Used carefully, this helps. Used carelessly, models trained mostly on other models' output can drift away from reality.
Everyday example
A pilot practising in a flight simulator
Trainee pilots spend hours in simulators before flying a real plane. The simulator can throw engine failures, storms and tricky landings at them, things that rarely happen in real flights and would be dangerous to arrange. The practice is not real, but it is realistic enough to build real skill. Synthetic data plays the simulator's role for AI.
Where the example breaks: A flight simulator is built from physics and checked by experts. Synthetic data is only as good as the generator that made it, and it can quietly miss rare real cases or copy the generator's own mistakes and biases.
How it actually works
Synthetic data ranges from simple rule-based fakes to data produced by generative AI models trained on real data.
Fully or partly synthetic
Fully synthetic data is entirely generated. Partly synthetic data keeps real records but replaces sensitive fields, such as names or exact incomes, with artificial values to protect privacy.
How it is made
Statistical models, simulators and generative AI such as GANs, diffusion models and language models. The generator learns the patterns of real data, then samples new examples from them.
What it is used for
Filling gaps, such as rare fraud cases; sharing data without exposing people; testing software; and training models where real labelled data is scarce. Every synthetic example comes with its label already known.
Training language models
Microsoft's phi-1 (2023), a 1.3 billion parameter code model, was trained partly on textbooks and exercises written by GPT-3.5. Many labs now use model-written examples in fine-tuning and distillation.
Model collapse
A 2024 Nature paper found that indiscriminate training, generation after generation, on model-made content causes irreversible defects in which the rare tails of the original data disappear. The usual defence is to keep real data in the mix and check synthetic data carefully.
- Early idea
- Statistician Donald Rubin proposed fully synthetic data in 1993, for census privacy
- Common uses
- Privacy-safe data sharing, rare-event training, software testing, simulation for robots and cars
- Close cousin
- Data augmentation: making new examples by transforming real ones, such as flipping images
- Main risk
- Copying the generator's errors and biases; model collapse if overused
Can synthetic data replace real data?
One view
It can carry much of the load
Generated data is cheap, private and unlimited, and models trained partly on it, such as phi-1, have done well. As real web text runs short, synthetic data looks like a way to keep improving models.
The other view
It cannot stand alone
Generated data can only reflect what the generator already knows. Research on model collapse shows that training repeatedly on model output loses rare cases and drifts from reality.
Most likely: Synthetic data works best as a supplement: mixed with real data, filtered and checked, rather than as a full replacement.
Questions people ask
What is synthetic data in simple terms?
Data that a computer made up instead of collecting from the real world, designed to have the same patterns as real data. It might be fake customer records, computer-generated images or AI-written practice questions.
Is synthetic data the same as fake data?
It is artificial, but not random. Good synthetic data is generated to match the statistics and structure of real data closely enough to be useful for training or testing, while not describing any real person or event.
Why do companies use synthetic data?
To get examples they cannot easily collect, such as rare fraud cases; to share or test with data that contains no real personal details; and to save the time and cost of gathering and labelling real data.
What is model collapse?
A problem where AI models trained repeatedly on content produced by other AI models get steadily worse. A 2024 Nature study found that the rare, unusual parts of the original data disappear first. Mixing in real data helps prevent it.