Skip to content
Article reader Listen + reading controls
LISTEN + READ YOUR WAY

Article reader

Preparing the reader…

0:00 0:00
Reading settings
Text size
100%

Synthetic Data Is a Claim About the World

Synthetic data is frequently presented as a practical escape from real-data constraints. When records are sensitive, scarce, imbalanced, expensive, or difficult to share, an organization can generate artificial observations that resemble the original population and continue developing models, testing systems, or publishing evidence.

That description is attractive and incomplete.

A synthetic dataset is not simply “fake data.” It is the output of a model that claims to preserve some properties of the world while suppressing or altering others. Its value depends on which properties survive, for which population, under which intended use, and with what privacy and disclosure risk.

The Federal Chief Data Officers Council’s 2024 request for information was therefore right to ask not only about applications, but about definitions, limitations, ethics, equity, and evidence. Federal synthetic-data practice should begin with a disciplined principle: every synthetic dataset is a bounded, testable claim.

The CDO Council’s RFI sought input from data leaders, scientists, technologists, stewards, and evidence-building experts to inform government best practices. The breadth of the request matters because synthetic data sits at the intersection of statistics, machine learning, privacy, security, records management, program evaluation, and operational decision-making.

No single fidelity score can govern that intersection.

Begin With the Intended Use

“Is the synthetic data accurate?” is the wrong first question. Accuracy is always relative to a task.

The same dataset may be appropriate for:

  • teaching analysts how to use a tool;
  • testing whether a pipeline handles the correct schema;
  • developing interface prototypes;
  • balancing classes during model training;
  • estimating population statistics;
  • evaluating a policy;
  • or supporting a consequential operational decision.

Each use requires different properties. A schema-valid training dataset may need realistic field formats but not preserve correlations. A dataset used for statistical research must preserve distributions and relationships relevant to the analysis. Data used to train a safety-impacting model must represent rare and consequential conditions, not merely average cases.

A federal synthetic-data artifact should therefore include an intended-use statement:

  1. Which task and decision will it support?
  2. Which real population, environment, and time period does it represent?
  3. Which statistical or semantic properties must be preserved?
  4. Which uses are prohibited or unsupported?
  5. Which evidence demonstrates suitability?
  6. Which changes would invalidate that evidence?

Without this statement, “high fidelity” becomes an invitation to reuse the data beyond its tested scope.

Privacy Is Not Automatic

Replacing direct identifiers or generating artificial rows does not guarantee privacy. A generative model may memorize rare records, reproduce sensitive combinations, permit membership inference, or expose attributes through correlations. Small populations and outliers can remain vulnerable even when every row is technically synthetic.

Privacy evaluation should match the threat model:

  • Can an attacker determine whether a person contributed to the source data?
  • Can the attacker infer a sensitive attribute using public auxiliary information?
  • Does the generator reproduce rare or unique records?
  • How does repeated release or querying compound disclosure risk?
  • Who can access the generator, parameters, source data, and outputs?
  • Are formal protections such as differential privacy used, and what utility cost follows?

Formal privacy guarantees can be valuable because they bound risk under stated assumptions. They do not make every downstream use safe. Governance still needs access control, release review, provenance, and monitoring for composition across datasets.

Synthetic data should be treated as a privacy control with evidence, not as a new data classification that is presumptively safe.

Fidelity Has Multiple Dimensions

Synthetic data can resemble source data in one way and diverge in another. A useful evaluation separates:

  • marginal fidelity: distributions of individual variables;
  • relational fidelity: correlations, dependencies, sequences, and causal structure;
  • tail fidelity: rare values and low-frequency events;
  • subgroup fidelity: behavior for demographic, geographic, clinical, or mission-relevant groups;
  • temporal fidelity: changes, seasonality, and ordering over time;
  • semantic fidelity: whether records remain meaningful under domain rules;
  • task utility: performance of a defined analysis or model trained or tested on the data;
  • privacy risk: resistance to reidentification, memorization, and inference.

Improving one dimension can harm another. Strong privacy protection may reduce tail fidelity. Matching aggregate distributions may conceal broken causal relationships. A generator may produce statistically plausible records that violate domain constraints.

The evaluation report should expose those tradeoffs rather than compress them into a composite score.

Synthetic Data Can Reproduce and Conceal Bias

If source data reflects unequal access, enforcement, labeling, or measurement, the generator can reproduce those patterns. It may also smooth away visible disparities, making the synthetic dataset appear more balanced while remaining unrepresentative of the real system.

Conversely, teams may intentionally alter distributions to create a more balanced training set. That can be useful, but it changes the dataset from a representation of observed history into a design intervention.

Both choices require documentation:

  • Which disparities were preserved, reduced, or amplified?
  • Did affected communities participate in defining acceptable representation?
  • Are historical inequities relevant evidence, bias to mitigate, or both?
  • How will a model trained on altered data behave when deployed into the original environment?
  • Does synthetic balancing improve model performance for a group or merely improve a fairness metric?

Equity cannot be inferred from the artificiality of the records. It must be evaluated in the downstream decision system.

Provenance Must Follow the Synthetic Record

Synthetic data needs a provenance model that records:

  • source datasets and their legal authority;
  • generator and software versions;
  • training and privacy parameters;
  • preprocessing and exclusions;
  • intended use and prohibited use;
  • evaluation methods and results;
  • known limitations;
  • release and access decisions;
  • and relationships to later derived datasets and models.

This lineage should travel with the artifact in machine-readable form. A dataset downloaded from a catalog without its generation context is easy to mistake for observed evidence.

Provenance also supports reproducibility. If a model’s performance changes, teams can distinguish changes in the learning algorithm from changes in the synthetic-data generator or source population.

Separate Development, Evaluation, and Evidence Data

One of the highest-risk practices is evaluating a model on synthetic data produced through assumptions similar to those used to train it. The process can create a closed epistemic loop: the generator and predictive model agree because they share a simplified representation of reality.

Where feasible, teams should separate:

  • synthetic data for development and experimentation;
  • real or independently generated data for evaluation;
  • and authoritative observed evidence for consequential validation.

When real evaluation data cannot be used, the limitation should narrow the deployment claim. Multiple generators, expert challenge, adversarial scenarios, simulation, and carefully controlled pilots may reduce uncertainty; they do not eliminate the missing evidence.

Synthetic data is particularly valuable for edge cases and simulation, but those cases should be anchored in domain knowledge and observed failure—not generated solely from what the model already believes is likely.

A Federal Synthetic-Data Registry

The government would benefit from a registry that connects synthetic datasets to their evidence and use. A registry entry could include:

  • responsible agency and steward;
  • source authority and sensitivity;
  • intended and prohibited uses;
  • generation method and version;
  • privacy claim and threat model;
  • fidelity and utility evidence;
  • affected populations and equity analysis;
  • expiration or review date;
  • known downstream models and publications;
  • and incident or misuse history.

The registry need not expose sensitive source details publicly. Internally, it would make synthetic data governable as a first-class asset rather than an informal workaround to access controls.

The Strategic Inference

Synthetic data can expand access, accelerate experimentation, improve testing, represent rare events, and support privacy-preserving analysis. Its strategic value lies precisely in allowing organizations to reason where direct observation is constrained.

That is also its epistemic risk. A generated dataset can look complete and precise while encoding an incomplete model of the world. It can hide uncertainty behind rows and columns that resemble ordinary evidence.

The correct question is not whether synthetic data is “realistic.” It is whether the specific properties needed for a specific use have been preserved with sufficient evidence, while privacy, bias, and misuse remain within acceptable bounds.

Used that way, synthetic data becomes a governed instrument of inference. Used as a generic substitute for reality, it becomes a source of confident error.

This essay was substantially revised in July 2026 to replace the original request-for-comment summary with an evidence-based governance analysis.

For related work on data provenance, knowledge infrastructure, and trustworthy AI operations, see my portfolio or connect on LinkedIn.

References

READER-NEUTRAL SUBSCRIPTION

Follow Field Notes via RSS.

Copy this address into the RSS reader you already use. New notes will appear there automatically—no account, email address, or tracking required.

Open the feed