What is synthetic data?
The short answer
Synthetic data is made by an algorithm that copies patterns from real data. The UK ICO says a model trained on the original generates it. Whether it's anonymous depends on whether personal information can be inferred. The more it copies real data, the more useful and revealing it gets.
The term shows up in AI training, privacy law and data licensing. I'm keeping each source's definition separate here.
How is synthetic data made?
The ICO describes two kinds. Partially synthetic data swaps out some variables of the original, such as location and admission time in a hospital dataset. It keeps the rest. Fully synthetic data swaps out all of them.
Either way, the generator needs real data first. So the original records are still processed once, even when the output holds none of them.
Is synthetic data anonymous?
The ICO says it depends on one thing. Can the personal information behind the synthetic set be inferred from the output? It adds that the more closely synthetic data mimics real data, the more useful it is. It is also more likely to reveal someone's details.
The ICO also notes that judging re-identification risk for synthetic data is still a young field. Its guidance is under review after the Data (Use and Access) Act. This is UK guidance, so your state's law may differ.
Where does synthetic data fit in AI training?
Troveo's guide lists four situations:
- edge cases
- privacy-sensitive domains
- labeling
- balancing underrepresented classes
It also names risks. One is model collapse, when models train heavily on model output. Another is a gap between generated and real-world texture.
Troveo sells real data, so it argues from a vendor's seat. Our real data vs synthetic data page sets that against Epoch AI's review of environments.
How does it connect to licensing company data?
Epoch AI describes using environments to generate data for supervised training, with successful trajectories as examples (Epoch AI). Licensed company records can feed the same kind of generation.
So a license can reach beyond the raw records. Ask whether it lets the buyer generate synthetic data from yours. Ask what the agreement says about that output. Our RL environment page shows one place such data gets used.
Start with the rights check, and talk to a privacy or commercial lawyer in your state before you sign.
Our only referral relationship is with micro1, an AI data company. It may pay us a fee if a company we introduce goes ahead with it. You pay nothing extra. How we are paid
Vaultfish helps you run the process. We are not lawyers and this is not legal advice. If you need counsel who knows this area, we can point you to one. Ask for a referral
