03 · What You Need to Know
What Transparent Synthetic-Data Reporting Should Include
Why must synthetic data be reported explicitly?
Readers need to understand the origin of the evidence supporting a study's findings.
Synthetic observations are generated through computational procedures rather than directly collected from the people, instruments, or events they may resemble.
For example, 500 artificial questionnaire records do not establish that 500 participants completed a survey.
If a manuscript describes synthetic records simply as participant responses, readers may incorrectly interpret the sample size, recruitment process, and statistical findings.
Explicit reporting helps prevent this confusion and allows the generation procedure to be evaluated as part of the research methodology.
The distinction between synthetic data and genuine research observations should therefore remain clear throughout the manuscript.
Where should synthetic-data use be reported?
Reporting may be necessary in several parts of a research paper, depending on how central the artificial data are to the study.
| Manuscript section |
Information to report |
| Title and abstract |
Identify simulation or synthetic-data methods when they define the study design or materially affect interpretation. |
| Introduction |
Explain the methodological reason for using synthetic data. |
| Methods |
Describe generation procedures, assumptions, source data, validation, and analytical use. |
| Results |
Distinguish findings obtained from synthetic data from those supported by empirical observations. |
| Discussion |
Explain limitations, generalizability, and implications for future empirical research. |
| Data availability statement |
Clarify whether synthetic datasets, generation code, or underlying empirical records are available. |
| AI disclosure statement |
Describe relevant AI assistance according to journal or institutional requirements. |
Not every manuscript requires extensive discussion in every section. Reporting should be proportionate to the role synthetic data play in the study.
How should researchers explain why synthetic data were used?
The manuscript should identify the scientific or methodological purpose of generating artificial observations.
For example, synthetic data may be used to test statistical software, evaluate an estimator, develop a prediction model, or reduce direct exposure of confidential participant records.
These purposes have different implications.
A dataset generated for software testing does not necessarily need to reproduce every characteristic of a real population.
By contrast, synthetic data intended to support prediction-model development may require careful evaluation of predictive relationships and subgroup representation.
The stated purpose should therefore guide both the generation procedure and the validation strategy.
What information about the generation method should be included?
Researchers should describe the procedure sufficiently for readers to understand how the artificial observations were produced.
Relevant details may include the statistical model, generative algorithm, software implementation, generation parameters, and source information.
If the dataset was produced through a simulation, the manuscript should describe the data-generating assumptions and relevant parameter values.
If a generative model was fitted to empirical records, the methods should explain the source dataset and training procedure at an appropriate level of detail.
If a language model generated structured observations, the report should identify relevant model information and the instructions or prompting procedure used.
The objective is methodological transparency, not merely naming the software.
Should the exact AI model be identified?
Generally, when the model materially affects the generation procedure and the information can be disclosed.
Different AI models may produce different synthetic records even when given similar instructions.
Commercial models may also change over time, affecting reproducibility.
Researchers should therefore consider documenting the provider, model identifier, available version information, generation date, relevant settings, and other consequential configuration details.
However, model names alone may be insufficient because generation also depends on prompts, source data, sampling settings, and subsequent processing.
Where exact model versions are unavailable, that limitation should be acknowledged rather than replaced with an unsupported claim of reproducibility.
Should prompts used to generate synthetic data be reported?
When prompts materially determine the generated records, they form part of the methodological procedure.
For example, a prompt instructing a language model to generate hypothetical student responses with particular demographic characteristics and relationships among variables may substantially influence the resulting dataset.
Researchers should preserve relevant instructions, prompt variants, and generation settings where feasible.
Long prompts may be provided through supplementary materials or a research repository.
However, prompts containing confidential information should not be disclosed without appropriate authorization.
Researchers should also recognize that preserving a prompt does not guarantee identical outputs from a nondeterministic or changing AI system.
How should source data be described?
If synthetic data were generated from an empirical reference dataset, the manuscript should describe that source sufficiently to evaluate the generation method.
Relevant information may include the original sample characteristics, variables used, collection context, and applicable restrictions.
However, researchers should not disclose confidential participant information merely to make the generation process more transparent.
Where source records cannot be shared, the manuscript should explain the restrictions and identify any available controlled verification arrangements.
If no empirical source dataset was used, the report should instead describe the theoretical assumptions or rules that informed generation.
This distinction helps readers understand whether the artificial records reflect observed patterns or researcher-specified conditions.
How should the number of synthetic observations be reported?
Artificial record counts must be distinguished from genuine participant or experimental sample sizes.
For example, a study may analyze 1,000 synthetic observations generated from a reference dataset containing 250 real participants.
The manuscript should not imply that 1,000 independent participants were recruited.
Researchers should identify the number of original empirical records where relevant, the number of generated records, and how each contributed to the analysis.
Generating additional artificial rows does not automatically increase the amount of independent empirical information.
This distinction is essential when interpreting standard errors, confidence intervals, statistical significance, and claims about population representativeness.
How should synthetic-data validation be reported?
Researchers should explain which properties were evaluated and why those properties matter for the intended application.
For data intended to resemble an empirical reference dataset, relevant assessments may include distributions, correlations, conditional relationships, rare cases, and subgroup characteristics.
For prediction-model development, researchers may report performance on appropriate independent empirical test data.
For software testing, validation may concern expected outputs, data constraints, and edge cases.
Privacy-related synthetic datasets may require disclosure-risk evaluation.
The process of validating synthetic data should be reported with enough detail for readers to assess both favorable and unfavorable findings.
Should researchers report unsuccessful generation attempts?
Material failures and consequential changes to the generation procedure should be reported when they affect interpretation or reproducibility.
For example, a researcher may test several generators before selecting one that adequately preserves a particular relationship.
Reporting only the successful model without explaining the selection procedure could give an overly favorable impression of the method.
However, not every exploratory trial necessarily requires extensive discussion in the main manuscript.
Researchers should distinguish preliminary development from prespecified evaluation and document the selection process appropriately.
Where practical, supplementary materials can provide additional details about alternative models, unsuccessful configurations, and sensitivity analyses.
How should statistical results from synthetic data be interpreted?
Statistical results should be interpreted in relation to the artificial data-generating process.
For example, a correlation observed in generated survey responses does not independently establish that the relationship exists among actual students.
If the generator was designed to reproduce a positive association, observing that association may largely reflect the generation assumptions.
Similarly, statistical significance calculated from synthetic observations should not automatically be treated as evidence from additional independent empirical participants.
Researchers should explain whether the analysis evaluates a method, examines model behavior, or supports a claim independently validated using genuine observations.
The conclusions should remain consistent with the available evidence.
How should privacy protections be reported?
When synthetic data are generated from sensitive source records, researchers should describe relevant privacy safeguards and their limitations.
These may include access controls, data minimization, privacy-preserving generation methods, or disclosure-risk evaluations.
If differential privacy is used, the mechanism and relevant privacy parameters should be reported where appropriate.
Researchers should avoid describing synthetic data as automatically anonymous.
The possibility of privacy leakage from synthetic datasets means that privacy claims require their own evidentiary basis.
Statistical fidelity does not establish privacy protection.
How should limitations be discussed?
Limitations should address the characteristics of the synthetic data that affect scientific interpretation.
For example, the generator may inadequately reproduce rare cases, subgroup patterns, nonlinear relationships, or characteristics of populations not represented in its reference data.
A language-model-based simulation may depend strongly on prompt wording or model version.
Researchers should explain how these limitations constrain the findings.
Generic statements that synthetic data may differ from real data are less useful than identifying the specific differences relevant to the study.
The discussion should also distinguish limitations of the generator from limitations of the empirical source dataset.
Can researchers claim that synthetic data reproduce real-world behavior?
Only to the extent supported by appropriate validation.
A synthetic dataset may reproduce selected statistical characteristics of a reference sample without representing the complete behavior of the target population.
For example, an AI-generated student survey may approximate overall response distributions while failing to capture differences across institutions or demographic groups.
Researchers should avoid interpreting numerical similarity as proof of behavioral equivalence.
Claims about real-world generalization require an evidentiary basis beyond the artificial records themselves.
The distinction is especially important when synthetic data are used to simulate human participants.
What should a data availability statement explain?
The data availability statement should identify which datasets and supporting materials can be accessed.
For example, researchers may make synthetic records and generation code publicly available while restricting access to the original empirical dataset.
Where source data are confidential, the statement should explain the relevant restrictions without disclosing protected information.
Researchers should also clarify whether the shared synthetic dataset is the exact dataset used in the reported analyses or an illustrative replacement.
Sharing synthetic records does not automatically permit independent reproduction of analyses originally conducted on restricted empirical observations.
The availability statement should accurately describe what can and cannot be verified using the released materials.
Does AI disclosure replace synthetic-data reporting?
No. A general statement that generative AI was used does not necessarily explain how artificial observations were generated or analyzed.
For example, a disclosure identifying an AI model does not establish which variables were generated, what assumptions were used, or whether empirical records were involved.
AI disclosure and methodological reporting serve related but distinct purposes.
The International Committee of Medical Journal Editors emphasizes transparency about AI assistance and the continuing responsibility of human authors.
Researchers should follow applicable publication requirements while ensuring that the methods section provides sufficient information to evaluate the synthetic-data procedure itself.
Are there universal reporting guidelines for synthetic data?
There is no single reporting checklist universally applicable to every synthetic-data study across all disciplines and purposes.
However, existing methodological and reporting frameworks can provide useful guidance.
For simulation studies evaluating statistical methods, Morris, White, and Crowther propose the ADEMP framework, which addresses aims, data-generating mechanisms, estimands, methods, and performance measures.
For prediction-model studies, TRIPOD+AI provides relevant reporting guidance.
FAIR data principles and established provenance frameworks can also inform documentation and data stewardship.
Researchers should select reporting guidance appropriate to their study rather than assume that one framework covers every synthetic-data application.
