What is Synthetic data?

Synthetic data is data that was made, not gathered. Instead of scraping a real web page or recording a real customer call, someone (or something) generates the example. Today that usually means one AI model writing text, code, math problems, conversations or images that another model then learns from.

How it works

The simplest recipe has three steps. A capable model generates a large batch of examples. A filter, a rule, a test or a second model checks them and throws out the bad ones. The survivors become training data for the next model or for a fine-tuning run.

Code is a good fit because you can check the answer. A model writes a function, a test suite runs it, and only code that passes is kept. Math works the same way: the final answer is either right or wrong. Where correctness can be verified, synthetic data can be high quality at huge scale.

Synthetic data also shows up outside training. Teams use it to test software without touching real customer records, and to fill gaps where real examples are rare, such as unusual fraud patterns or edge cases in a safety evaluation.

Why it matters

Real text is finite, and much of the best of it is already used. Synthetic data is one way labs keep feeding models. It is also a privacy tool: fake records that look statistically like real ones can let you build and test without exposing anyone's details.

It is closely tied to distillation, where a smaller model learns from a larger model's outputs, and to post-training, where models are often tuned on generated examples of good answers.

The risk: model collapse

The known danger is training models on their own output, round after round. Researchers call the result model collapse: performance degrades with each cycle, outputs get more uniform, and the model gradually forgets the variety of the real data. Studies describe it as a risk of recursive loops with too little fresh real data, not a flaw in synthetic data itself. That is why careful pipelines keep real data in the mix and filter hard.

A simple example

Say you want a customer support bot that handles refund requests. You have 50 real transcripts. You ask a strong model to write 5,000 varied refund conversations in your style, including angry customers and vague requests, then you review a sample and discard the bad ones. You now have far more material to fine-tune on, though it is only as good as the model and the checks behind it.

What to remember

Synthetic data is a tool, not a shortcut. It works best when you can verify the output and keep real examples in the loop. Treat any claim that a model was trained mostly on synthetic data as a question about the checks, not a guarantee of quality.