03 · What You Need to Know
What Makes Synthetic-Data Use Scientifically and Ethically Defensible?
What does legitimate synthetic-data use mean?
Legitimate synthetic-data use involves generating or analyzing artificial observations for a purpose that can appropriately be addressed using such data.
Synthetic datasets may reproduce selected properties of existing empirical records or represent hypothetical conditions specified through a statistical or computational model.
The key consideration is whether the research question concerns something the artificial dataset can actually establish.
For example, synthetic data may be suitable for testing whether a statistical program handles missing values correctly. They may also support a simulation study examining estimator performance under specified assumptions.
However, the same records cannot independently establish what actual participants experienced or what happened in an experiment that was never conducted.
The legitimacy of the method therefore depends on a defensible relationship between the data-generation process and the intended scientific inference.
Which research activities commonly justify synthetic data?
Several research activities can benefit from artificial observations, although each requires different forms of validation.
| Research activity |
Legitimate role of synthetic data |
Important limitation |
| Statistical simulation |
Evaluating analytical methods under specified data-generating conditions |
Findings depend on the simulation assumptions. |
| Research software testing |
Checking data import, processing, calculations, and error handling |
Successful testing does not guarantee performance on every real dataset. |
| Prediction-model development |
Supplementing model development or testing selected scenarios |
Artificial data may not adequately represent real-world distributions. |
| Privacy-related data sharing |
Reducing direct exposure of original participant-level records |
Generated data may still create disclosure risks. |
| Methodological pilot testing |
Testing workflows, coding procedures, and analytical assumptions |
Simulation does not necessarily replace empirical validation. |
| Teaching and training |
Providing practice datasets with known or controlled characteristics |
Artificial observations must not be mistaken for empirical findings. |
The presence of a legitimate application does not mean every synthetic-data generation method is suitable for it.
Researchers must evaluate the generated records according to the requirements of the particular task.
Is synthetic-data generation legitimate when no real dataset exists?
Yes, for some purposes.
Researchers may generate artificial observations from mathematical assumptions, predefined rules, or hypothetical scenarios without using an existing empirical dataset.
For example, a statistical simulation may generate repeated samples under a specified regression model to evaluate estimator bias and confidence interval coverage.
Such a study does not need to claim that the simulated observations came from actual participants.
Its scientific contribution concerns how the analytical method behaves under the defined conditions.
However, a synthetic dataset generated without empirical reference data may be unsuitable for claims about the characteristics of a particular real population.
Researchers should distinguish methodological evidence from empirical evidence.
Does synthetic-data use require ethical approval?
Not every project using synthetic data requires human-participant ethics approval.
For example, a simulation based entirely on hypothetical parameters and containing no identifiable human information may fall outside the scope of human-subjects review in some jurisdictions.
However, synthetic data generated from confidential participant records may involve processing activities governed by an existing ethics approval, data-use agreement, or privacy law.
The fact that the final output is artificial does not necessarily remove obligations associated with the original data.
Researchers should consult the responsible institutional review body when the applicability of ethical requirements is uncertain.
They should not assume that the term synthetic automatically exempts a project from research ethics or data-protection requirements.
How does legitimate synthetic data differ from fabricated research data?
The United States Office of Research Integrity (ORI) defines fabrication as making up data or results and recording or reporting them.
In research-misconduct contexts, the relevant concern is the false representation of invented observations as genuine research evidence.
Synthetic data generated for an explicitly identified simulation or software-testing activity are different from invented observations falsely presented as empirical findings.
Legitimate synthetic data
Artificial records used for a justified purpose, with their origin, generation process, and limitations accurately represented.
Fabricated empirical data
Invented observations or results falsely recorded or reported as though they were genuinely obtained through research.
The distinction between synthetic and fabricated data therefore concerns representation and scientific purpose, not simply whether a computer generated the values.
Is transparent labeling sufficient to make synthetic data legitimate?
No. Labeling is necessary to prevent confusion about provenance, but methodological validity must be established separately.
For example, a researcher may clearly state that 500 survey responses were generated by a language model.
That disclosure does not establish that the responses accurately represent the attitudes of actual university students.
The artificial records may still be useful for software testing or methodological exploration, but population-level conclusions would require additional evidence.
Researchers should therefore distinguish truthful reporting from the validity of the resulting inference.
A study can be transparent about its artificial data while still making scientifically unsupported claims.
How should researchers validate synthetic data?
Validation should be tailored to the intended use rather than based on a single general measure of realism.
For synthetic data intended to approximate an empirical dataset, researchers may examine distributions, relationships among variables, subgroup characteristics, and downstream analytical performance.
For simulation studies, the priority may be verifying that the computational implementation matches the specified mathematical assumptions.
For software testing, researchers may deliberately include invalid entries, missing values, and unusual combinations to test whether a program behaves correctly.
Where synthetic data are derived from sensitive records, privacy evaluation may also be necessary.
The process of validating synthetic data should therefore establish fitness for purpose rather than superficial similarity alone.
Can synthetic data legitimately support published scientific findings?
Yes, when the findings concern questions that the artificial data can appropriately address.
For example, a simulation study may establish how a statistical method performs under specified assumptions.
A software-engineering study may demonstrate that a data-processing pipeline correctly handles defined test cases.
A methodological study may compare different synthetic-data generation techniques using prespecified evaluation criteria.
These are legitimate scientific contributions even though the underlying observations were artificially generated.
However, researchers should not present simulated outcomes as independent evidence that a real-world phenomenon occurred.
The evidentiary scope of the findings must remain consistent with the research design.
Can synthetic data be used when real participant data are unavailable?
Potentially, but their suitability depends on the research question.
If the objective is to test an analytical procedure, synthetic data may be sufficient.
If the objective is to investigate actual participant experiences, attitudes, or behavior, synthetic records generally cannot provide equivalent empirical evidence.
For example, AI-generated responses from fictional teachers may help test a coding framework, but they cannot establish what real teachers experienced when implementing educational technology.
The question of whether synthetic data can replace real participant data requires distinguishing methodological substitution from empirical substitution.
Recruitment difficulty alone does not justify presenting artificial respondents as genuine participants.
Can synthetic data improve prediction-model development?
Under some circumstances, synthetic data may support model development, particularly when genuine observations are limited or imbalanced.
However, generated records may reproduce biases or fail to preserve rare cases and important subgroup relationships.
A prediction model that performs well on synthetic data may not perform equally well on actual observations.
Researchers should therefore evaluate real-world performance using appropriate independent empirical data when that is the intended claim.
They should also consider whether synthetic training and testing datasets share generation artifacts that produce misleading performance estimates.
The legitimacy of synthetic data for prediction-model development depends on the evaluation design and intended deployment context.
Can synthetic data legitimately protect research participant privacy?
Synthetic data may reduce direct exposure of original participant records, but they do not automatically guarantee anonymity.
Some generation methods can reproduce distinctive combinations of characteristics or reveal information about individuals represented in the source data.
Privacy risk depends on the generation model, source records, release conditions, and information available to potential recipients.
Researchers should therefore assess disclosure risks rather than assume that artificial generation is sufficient protection.
Where formal privacy guarantees are required, appropriately designed mechanisms such as differential privacy may be relevant.
The question of whether synthetic data protect participant privacy must be evaluated separately from their statistical usefulness.
What makes synthetic-data reporting scientifically adequate?
Researchers should describe why synthetic data were generated and how they contributed to the study.
Relevant details may include the generation method, source data where applicable, model or software version, parameters, assumptions, validation procedures, and limitations.
When empirical and artificial records coexist, their origins and analytical roles should remain distinguishable.
Researchers should also follow applicable journal and institutional requirements concerning AI assistance.
A manuscript should not report an artificial sample size in a way that implies the corresponding number of human participants or experimental observations.
The issue of reporting synthetic data in research papers is therefore central to preserving accurate scientific interpretation.
When does synthetic-data use become methodologically inappropriate?
Synthetic-data use becomes inappropriate when the artificial observations cannot support the intended inference or when their limitations are concealed.
For example, generated survey responses should not be treated as evidence of population attitudes merely because their distribution resembles previous studies.
Similarly, artificial laboratory measurements cannot establish that an experiment produced the reported results.
Even when the dataset is accurately labeled, methodological overreach can produce unsupported conclusions.
Researchers should distinguish an invalid inference from formal research misconduct. Not every methodological weakness constitutes fabrication or falsification.
However, knowingly misrepresenting artificial observations as genuine evidence may raise serious misconduct concerns.
Who decides whether synthetic-data use is acceptable?
Researchers are responsible for selecting and justifying methods appropriate to their research questions.
However, institutions, ethics committees, funders, data custodians, and journals may impose additional requirements.
For example, a data-use agreement may restrict the processing of confidential source records, even when the intended output is synthetic.
A journal may require detailed disclosure of AI-assisted data generation or validation.
Researchers should therefore evaluate both methodological suitability and compliance with applicable governance requirements.
Scientific legitimacy is not established solely by the researcher's preference for a particular tool.
