---
title: "Synthetic data: definition and example – Carl-Uwe Würfel"
description: "Synthetic data explained: how it is generated, where it helps in testing, training, and evals, why synthetic doesn't mean anonymous, and what model collapse is."
url: https://carl-wuerfel.de/en/glossary/synthetische-daten
author: Carl-Uwe Würfel
---

1. [Home](https://carl-wuerfel.de/en)
  2. [Knowledge](https://carl-wuerfel.de/en/knowledge)
  3. [Glossary](https://carl-wuerfel.de/en/glossary)
  4. Synthetic data

# Synthetic data

Synthetic data is artificially generated data that reproduces the structure and statistical properties of real data without containing real individual records.

## On this page

  1. A generator learns the patterns and creates new cases
  2. Synthetic doesn't automatically mean anonymous
  3. Differential privacy makes the protection measurable
  4. It is used in testing, training, and evals
  5. Whatever is missing from the original is missing from the copy
  6. What good implementation looks like
  7. Where the term stops working
  8. Frequently asked questions
  9. Sources

Synthetic doesn't automatically mean anonymous.

_Figure: A decision diagram using the example of a utility that generates synthetic customer records with the same fields as its real data. The generator learns from real customer data; that training is already processing of personal data and needs a legal basis. What applies depends on the use: in testing the new customer portal, developers see no real customer data, but acceptance testing before launch still uses real data in a protected environment. Before sharing with a vendor, the team checks whether real records reappear almost unchanged. When training a model, part of the real data stays in the mix. Below: special cases such as legacy contracts are missing from the synthetic set._

Synthetic data looks like the real thing. A synthetic customer record has a name, address, rate plan, and payment history, but the customer doesn't exist. The US National Institute of Standards and Technology (NIST) describes it as data with the same schema as the original that tries to keep its properties, such as correlations between attributes, while consisting of fake individuals.

Most synthetic data comes from a model that has learned the patterns of a real dataset. There are also rule-based generators that assemble test data from fixed specifications, and language models that write text or test questions for other models.

For executives, synthetic data is a way to work with data that would otherwise have to stay in-house. It has two catches: synthetic doesn't automatically mean anonymous, and whatever is missing from the original is missing from the copy too.

Example

A utility launches a new customer portal. The test environment gets synthetic customer records with the same fields as the real ones, so developers and the vendor never see real customer data. Beforehand, the team checks whether individual real records reappear almost unchanged. Acceptance testing before launch still uses real data in a protected environment, because special cases such as legacy contracts are missing from the synthetic set.

Chapter 1 of 7

## A generator learns the patterns and creates new cases.

It starts with the purpose. For a software test, a rule-based generator that assembles valid customer IDs, addresses, and contract data is often enough. If the data needs to show real relationships, such as which rate plans go with which payment behavior, a model has to learn them from real records.

The team then selects and cleans the source data, trains the generator, produces new records, and checks them. The best-known type is the generative adversarial network, introduced by Ian Goodfellow and colleagues in 2014: one network generates data, a second tries to tell it apart from real data, and both improve in the process. For text, language models now play this role.

The final check answers two questions: whether the data serves its purpose, and whether it reveals anything about real people.

Chapter 2 of 7

## Synthetic doesn't automatically mean anonymous.

Under Recital 26, the GDPR does not apply to anonymous information that relates to no identified or identifiable person. The test is whether any means reasonably likely to be used could identify someone. Pseudonymized data, where names are merely replaced by codes, remains personal data.

Synthetic data may or may not meet that test. A generator can memorize individual records and reproduce them almost unchanged, especially unusual ones. In experiments published in 2022, Theresa Stadler and colleagues found that synthetic data either failed to prevent attacks on individuals or lost its usefulness. In December 2024, the European Data Protection Board stated that AI models trained on personal data cannot in all cases be considered anonymous.

Training the generator on customer data is itself processing under Article 4(2) GDPR and needs a legal basis under Article 6.

Chapter 3 of 7

## Differential privacy makes the protection measurable.

Differential privacy is a mathematical method that limits how much any single record can affect the result. A parameter, usually called epsilon, sets that limit: the smaller the value, the stronger the protection and the less accurate the result.

In guidelines published in March 2025, NIST recommends generating synthetic data with such methods when it needs to protect people reliably. According to NIST, techniques without differential privacy offer only informal guarantees, and there are reports of attacks that recovered the original data from such datasets. Even with differential privacy, outliers can remain exposed if epsilon is set generously.

The protection costs accuracy, and not evenly. NIST notes that synthetic data can become less accurate for small subgroups, which leads to systematic bias. Analyses of rare customer groups suffer most.

Chapter 4 of 7

## It is used in testing, training, and evals.

In software development, synthetic data replaces real customer data in test and development environments. Migrations and load tests then run without customer data ending up with developers or vendors.

In training, it supplements scarce data. In 2023, Microsoft researchers introduced phi-1, a model trained on 6 billion [tokens](https://carl-wuerfel.de/en/glossary/token) of selected web text plus 1 billion tokens of synthetic textbooks and exercises written by GPT-3.5. Rare cases, such as defect patterns in quality control, can be multiplied the same way, provided they appear in the source data at all.

For evals, Anthropic's documentation suggests having a model generate more test cases from a baseline set of examples. The baseline consists of real cases, and a person reviews a sample of the generated ones.

Chapter 5 of 7

## Whatever is missing from the original is missing from the copy.

A generator reproduces what it has learned. It carries over the biases in the source data and doesn't reliably invent cases that never occurred there. Enlarging a loan dataset with few self-employed applicants synthetically yields more rows, not a better picture of the self-employed.

The limit shows most clearly when models are trained on the output of models. In a 2024 Nature paper, Ilia Shumailov and colleagues showed that indiscriminate use of model-generated content in training causes irreversible defects: the tails of the original distribution, meaning the rare cases, disappear. They called this model collapse. In their experiments, keeping part of the original data in every round made the degradation much smaller.

Synthetic data supplements real data rather than replacing it. Acceptance testing and measuring results run against real data.

Chapter 6 of 7

## What good implementation looks like.

It starts with a defined purpose and the question of whether a rule-based generator is enough. Many software tests need no learned model and therefore no training on customer data.

If the generator learns from real data, the legal basis is settled before training starts. Utility is measured against the original: the same analysis on real and synthetic data should give a similar picture. The privacy check looks for records that nearly match real ones and for identifiable outliers, and its result is documented.

Synthetic records are labeled so nobody mistakes them for real ones or slips them into training unnoticed. Before an application goes live, it passes acceptance testing with real data in a protected environment.

Most of the cost goes into selecting and cleaning the source data and into the checks afterward, rarely into the generator itself.

Chapter 7 of 7

## Where the term stops working

Synthetic data does not replace measuring reality: whatever is missing or skewed in the source data is missing or skewed in the copy. Anonymized data consists of altered real records, pseudonymized data stays personal, and simulations follow fixed rules instead of learned patterns. Data augmentation, creating variants of existing examples, is a related but narrower technique.

## Frequently asked questions

What is synthetic data?

Synthetic data is artificially generated data that reproduces the structure and statistical properties of real data. It contains no real individual records but can be analyzed with the same tools.

How is synthetic data generated?

Usually a model, such as a generative adversarial network or, for text, a language model, learns the patterns of a real dataset and generates new records. Simple software tests can use rule-based generators that need no source data.

Is synthetic data GDPR compliant?

Synthetic data falls outside the GDPR only if no one can be identified from it, and that has to be checked case by case. Training the generator on personal data is itself processing and needs a legal basis.

What is the difference between synthetic and anonymized data?

Anonymized data consists of real records stripped of the personal link; synthetic data is newly generated. Both carry residual risk: anonymized data through linkage with other sources, synthetic data through records the generator memorized.

Can AI be trained on synthetic data?

Yes. As a supplement to real data it is common, as with the phi-1 model and its synthetic textbooks. Training models over many rounds only on model output risks model collapse.

What is model collapse?

Model collapse is the loss of quality in models that are repeatedly trained on the output of other models. Shumailov and colleagues showed in 2024 that rare cases in particular disappear from the data.

What is synthetic data used for?

Synthetic data is used mainly for software testing without real data, as a supplement when training AI models, and for test sets in [evals](https://carl-wuerfel.de/en/glossary/evals). It also lets companies share data with vendors while real data stays in-house.

## Further reading

  * [**AI glossary**](https://carl-wuerfel.de/en/glossary)
  * [**AI governance for mid-sized companies**](https://carl-wuerfel.de/en/ai-governance)
  * [**Data quality**](https://carl-wuerfel.de/en/glossary/datenqualitaet)
  * [**Evals (LLM evaluation)**](https://carl-wuerfel.de/en/glossary/evals)
  * [**Data governance**](https://carl-wuerfel.de/en/glossary/data-governance)
  * [ Term**Fine-Tuning**](https://carl-wuerfel.de/en/glossary/fine-tuning)
  * [ Term**Machine learning**](https://carl-wuerfel.de/en/glossary/machine-learning)

## Sources

  1. [Goodfellow et al.: Generative Adversarial Networks, arXiv:1406.2661, 2014](https://arxiv.org/abs/1406.2661)
  2. [European Union: Regulation (EU) 2016/679 (General Data Protection Regulation), Recital 26, Articles 4 and 6, 2016](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32016R0679)
  3. [European Data Protection Board: Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models, December 18, 2024](https://www.edpb.europa.eu/our-work-tools/our-documents/opinion-board-art-64/opinion-282024-certain-data-protection-aspects_en)
  4. [NIST: Guidelines for Evaluating Differential Privacy Guarantees, SP 800-226, March 2025](https://doi.org/10.6028/NIST.SP.800-226)
  5. [Stadler, Oprisanu, Troncoso: Synthetic Data, Anonymisation Groundhog Day, USENIX Security 2022, arXiv:2011.07018](https://arxiv.org/abs/2011.07018)
  6. [Gunasekar et al.: Textbooks Are All You Need, arXiv:2306.11644, 2023](https://arxiv.org/abs/2306.11644)
  7. [Shumailov et al.: AI models collapse when trained on recursively generated data, Nature 631, July 24, 2024](https://doi.org/10.1038/s41586-024-07566-y)
  8. [Anthropic: Define success criteria and build evaluations, documentation, as of November 2026](https://platform.claude.com/docs/en/test-and-evaluate/develop-tests)
