Jargon:Synthetic data
1 min read
Say this at the party
“Data made up by a computer instead of collected from the real world. AI companies now use it to train AI, because real data is running short.”
In plain English
To learn, AI needs examples. Real examples (books, photos, medical records) can be scarce, private or expensive.
So people generate fake but realistic ones instead. Sometimes a program makes them, and these days it's often another AI writing practice questions, stories or code.
Done carefully, it fills gaps. Done carelessly, the AI ends up learning from its own echoes and gets worse.
Where it came from
The idea predates modern AI. In 1993, statistician Donald Rubin proposed fully synthetic census data, so researchers could study it without exposing real households. In 2023, Microsoft's phi-1 paper showed a small model trained partly on AI-written textbooks could code surprisingly well.
See it in action
A hospital wants to build a tool that spots rare illnesses, but can't share patient files. It creates realistic made-up records that keep the patterns without belonging to any real person.
What people get wrong
- Fake doesn't mean useless. Good synthetic data is checked and mixed with real data, and it can teach a model a lot.
- It isn't endlessly safe to feed AI on AI output. A 2024 Nature paper found that training on model-made content without care can cause "model collapse", where rare details fade away.
Go deeper
- ArticleWikipediaSynthetic dataOverview and history, including Rubin's 1993 census idea.
- PaperGunasekar et al., Microsoft, 2023Textbooks Are All You NeedA small coding model trained partly on textbooks written by GPT-3.5.
- PaperShumailov et al., Nature, 2024AI models collapse when trained on recursively generated dataThe warning: AI trained carelessly on AI output loses the rare stuff.
Related jargon
Drafted with AI. The AI Council (Claude, ChatGPT and Gemini) gave counsel. Phin signed it off.