# Synthetic Data

Synthetic data consists of artificially generated records or examples created to support a defined testing, analysis, or model-development purpose.

- Type: artifact
- Domain: Engineering & architecture
- Level: Intermediate
- Last reviewed: 2026-10-06
- Original: [Soroush’s Product Wiki](<https://wiki.soroush.uk/topic/synthetic-data/>)

## What it means

Synthetic data is produced by rules, simulations, statistical models, or generative models instead of directly recording the target events. It may imitate selected properties of a real dataset or represent deliberately constructed scenarios. It differs from simply sampling real records or removing identifiers. Artificial origin does not establish realism, usefulness, or privacy protection.

## Why it matters

Teams can use synthetic examples to exercise rare workflows, prototype with restricted datasets, or supplement model-development material. The value depends on which properties are preserved: data adequate for testing a form may be unsuitable for estimating behavior or evaluating a predictive model.

## When to use it

Use synthetic data when its purpose is explicit and fitness for that purpose can be tested. It can support controlled edge cases and software testing. Do not treat simulated users or generated responses as direct evidence of real customer needs, preferences, or market demand.

## How it works

1. Define the intended use and the relationships, constraints, and edge cases the generated data must preserve.
2. Select a generation method and document its inputs, assumptions, permissions, and version.
3. Check schema validity, internal consistency, duplicates, and coverage of required scenarios.
4. Evaluate task-specific utility against appropriate real evidence and assess disclosure risk separately.
5. Label the dataset as synthetic, record unsuitable uses, and retain independent real-world validation for product conclusions.

## A practical example

A subscription team generates accounts with expired trials, overlapping entitlements, and unusual billing dates to test an AI support workflow. The examples reveal handling errors in rare states. The team uses real, appropriately governed support cases to estimate production quality because the generated mix does not represent actual customer frequency.

## Where it breaks

Synthetic data can omit rare relationships, reproduce generator biases, or resemble source records closely enough to expose information. Preserving broad distributions does not establish usefulness for every task. A large generated dataset does not create independent evidence when all records reflect the same flawed assumptions.

## Common mistakes

- Assuming synthetic means anonymous or safe to publish.
- Using generated respondents as a substitute for customer research.
- Measuring a model only on data produced by the same model or assumptions.
- Reusing test fixtures for statistical conclusions without validating representativeness.

## Sources

1. [Synthetic data policy](<https://www.ons.gov.uk/aboutus/transparencyandgovernance/datastrategy/datapolicies/syntheticdatapolicy>) — Office for National Statistics · Government guidance
2. [SDNist: Synthetic Data Report Tool](<https://www.nist.gov/services-resources/software/sdnist-synthetic-data-report-tool>) — National Institute of Standards and Technology · Government technical documentation

## Related topics

- [Dataset Card](<https://wiki.soroush.uk/topic/dataset-card/>)
- [Model Evaluation](<https://wiki.soroush.uk/topic/model-evaluation/>)
- [Research Validity and Reliability](<https://wiki.soroush.uk/topic/research-validity-and-reliability/>)
