Synthetic data

In plain English

Synthetic data is a flight simulator for AI. Pilots practice realistic scenarios, including rare emergencies, without risking a real plane. Synthetic data lets a model learn from realistic records, including rare cases, without exposing the real people behind the original data.

A generator, often itself a machine learning model, studies a real dataset and produces new artificial records. NIST describes a synthetic dataset as having the same structure as the original but consisting of generated data about "fake" individuals (NIST SP 800-226). Good synthetic data keeps useful patterns, such as how age relates to blood pressure. It should not reproduce any actual person.

Two warnings come with it. If the generator copies real records too closely, the "fake" data can reveal real people. And synthetic data inherits the gaps and biases of the data it learned from.

Example

Ashby Vale Health, a fictional hospital group, wants to train a sepsis model but has few records for patients under 18. It generates synthetic pediatric records from its real data and runs three checks before use:

The privacy check flags near copies of several rare cases. The team regenerates the data with a differentially private method. Its datasheet records that the pediatric data is synthetic, how it was made and which checks it passed. Final validation still uses real patient data.

Why it matters

• Privacy. Training a generator on personal data is processing under the GDPR (Art. 4(2)). The output falls outside the GDPR only if people are no longer identifiable (Recital 26). NIST warns that synthetic data made without differential privacy gives only informal protection, and attacks have revealed original data.
• Bias detection under the EU AI Act. The 2026 Digital Omnibus moved the special-category data rule to a new Art. 4a. High-risk providers may process such data for bias detection only if synthetic or anonymized data would not work (Art. 4a(1)(a)).
• Training data transparency. GPAI model providers must publish a training content summary (Art. 53(1)(d)). The Commission's template lists synthetic data among the source types to describe.
• Not the same as synthetic content. Art. 50(2) of the AI Act requires generative systems to mark synthetic audio, image, video or text outputs. That rule concerns outputs shown to people, not training data.

How the AIGP exam tests it

Synthetic data appears with the other privacy-enhancing techniques in lesson III.B.5. It also appears as a minimization tool in lesson II.A.2. The exam gives each technique one job.

• The tell. Real examples are too scarce, or a rare case needs coverage: synthetic data. Raw data cannot leave its source: federated learning. Published results must not reveal any individual: differential privacy.
• Trap 1: "synthetic data can never reveal real people." False. Poorly made synthetic data can replicate real records, so it needs leakage testing.
• Trap 2: synthetic data as a quick fix. If a dataset underrepresents a group, first find out why through provenance and lineage. Synthetic records built on a wrong assumption encode that assumption.
• Trap 3: skipping real-world validation. A model trained on synthetic data must still be validated on real data from its intended setting.