03 · What You Need to Know
How Synthetic Data Differ From Research Data Fabrication
What are synthetic data?
Synthetic data are artificially generated records produced through computational procedures rather than directly obtained from the real-world events or entities they may resemble.
They may be generated using statistical simulation, probabilistic models, generative adversarial networks, diffusion models, large language models, or other methods.
Some synthetic datasets are designed to reproduce selected statistical properties of empirical records. Others represent hypothetical scenarios or controlled mathematical assumptions.
For example, a researcher may generate artificial student engagement responses to test whether a statistical script correctly handles reverse-coded questionnaire items.
Such records are not genuine student responses, but they can be useful for evaluating the analytical procedure.
The legitimacy of AI-generated synthetic research data depends on the purpose, method, documentation, and interpretation.
What is research data fabrication?
The United States Office of Research Integrity (ORI) defines fabrication as making up data or results and recording or reporting them.
Within the research-misconduct framework, this concerns invented observations or results being represented as genuine research evidence.
For example, a researcher who creates fictional participant responses and reports them as completed questionnaires misrepresents the data collection process.
Likewise, reporting invented laboratory measurements as results from experiments that never occurred constitutes fabrication.
The use of generative AI does not change the underlying definition.
Fabrication can occur through manual entry, spreadsheets, statistical software, or AI-assisted workflows.
Formal findings of research misconduct nevertheless require the applicable evidentiary and culpability standards to be satisfied.
What is the main difference between synthetic and fabricated data?
| Dimension |
Synthetic data |
Fabricated research data |
| Origin |
Artificially generated. |
Made up rather than genuinely obtained as claimed. |
| Representation |
Identified as artificial or simulated. |
Falsely recorded or reported as genuine research evidence. |
| Typical purpose |
Simulation, software testing, methodological development, or other justified applications. |
Creating a false research record or unsupported appearance of empirical findings. |
| Participant status |
Generated respondents are not counted as actual human participants. |
Fictional respondents may be falsely reported as real participants. |
| Scientific claims |
Limited by the generation method and its validation. |
Claims are misrepresented as supported by observations that were not obtained. |
| Research integrity |
Potentially legitimate when appropriately designed and reported. |
Falls within fabrication definitions when the applicable misconduct criteria are met. |
The distinction does not depend on how convincing the values appear.
A synthetic dataset can be statistically realistic without containing any new empirical observations. A fabricated dataset can be statistically implausible or highly convincing while remaining falsely represented.
Can the same dataset be synthetic in one study and fabricated in another?
Yes. The same artificial records may be used legitimately or misrepresented, depending on their role in the research.
Suppose a researcher generates 200 hypothetical questionnaire responses.
Using those responses to test a data-processing pipeline may be legitimate when the records are identified as synthetic.
However, reporting that 200 actual students completed the questionnaire would falsely represent the empirical sample.
The numerical content has not changed. What has changed is the claim made about its origin.
This illustrates why research integrity cannot be evaluated solely by inspecting the dataset.
Researchers must also examine the documented methods and how the observations are described in the research record.
Does generating synthetic data constitute fabrication under ORI's definition?
Not merely because artificial data were created.
Statistical simulation, methodological testing, and other legitimate research activities routinely involve generating hypothetical observations.
ORI's definition concerns making up data or results and recording or reporting them in the context of research misconduct.
Clearly identified artificial data used for an appropriate methodological purpose should not be equated with invented empirical observations falsely presented as genuine.
However, labeling a dataset synthetic does not automatically resolve every research-integrity concern.
A study may still make unsupported claims, misrepresent its analytical methods, or violate applicable research requirements.
The relevant evaluation must consider the complete research process.
Does disclosure automatically make synthetic data legitimate?
No. Transparency is necessary, but methodological suitability is a separate requirement.
For example, a researcher might disclose that an entire survey dataset was generated by AI and then claim that the results establish actual university students' attitudes toward online learning.
The disclosure identifies the data's origin, but it does not independently establish that the generated responses accurately represent the target population.
The conclusions would require appropriate empirical or other independent validation.
Researchers should distinguish two questions: whether the data were represented honestly and whether the data can support the intended scientific inference.
A study may satisfy the first requirement while failing the second.
Can synthetic data be used to increase a research sample size?
Not by treating generated records as additional genuine participants or experimental observations.
Suppose a survey contains 180 authentic responses. Generating another 120 artificial records does not mean that 300 people participated.
If all 300 records are reported as genuine responses, the empirical sample has been misrepresented.
Artificial data may be used in particular model-based or simulation approaches, but their analytical role must be justified and their provenance preserved.
Researchers should also distinguish synthetic-data generation from legitimate missing-data imputation.
Imputation estimates missing information under specified assumptions. It does not establish that additional participants were recruited or that missing measurements were directly observed.
Can synthetic data produce statistically significant findings?
Yes. Statistical procedures can produce coefficients, confidence intervals, p-values, and other numerical results when applied to artificial datasets.
However, those results must be interpreted in relation to the generation process.
A statistically significant association in a synthetic dataset does not independently demonstrate that the same relationship exists in a real population.
For example, if artificial student engagement scores were generated using assumptions that already specify a positive relationship with instructor support, observing that relationship in the resulting data would not constitute independent empirical confirmation.
The analysis may still be valuable for evaluating an estimator or demonstrating statistical behavior.
Researchers should avoid interpreting model-generated patterns as discoveries about real-world participants without appropriate supporting evidence.
Can synthetic data be used legitimately in published research?
Yes. Synthetic data can support publishable methodological research when the design and reporting are appropriate.
For example, a study might evaluate how a statistical estimator performs under different missing-data mechanisms using repeated simulated datasets.
Another study might examine whether synthetic training data improve the performance of a prediction model when evaluated against an independent empirical test set.
These designs can generate meaningful scientific findings about methods and systems.
However, the conclusions must remain within the scope of the evidence.
Researchers should report the artificial data-generation process, assumptions, validation procedures, and limitations.
The question of reporting synthetic data in research papers is therefore central to both transparency and methodological interpretation.
Can synthetic data be generated from real participant records?
Yes. Some synthetic-data methods learn patterns from empirical datasets and generate artificial records intended to preserve selected statistical characteristics.
Such methods may reduce the need to share original participant-level observations.
However, synthetic generation does not automatically guarantee privacy.
A model may reproduce distinctive records or reveal information about individuals represented in the source dataset.
Researchers should therefore evaluate disclosure risks when generating synthetic data from confidential or identifiable information.
The possibility of privacy leakage from synthetic data is separate from whether the data are being represented honestly.
A dataset can be legitimately synthetic while still presenting unacceptable privacy risks.
What if a researcher mixes synthetic and real observations?
Combining artificial and empirical information may be appropriate for certain methodological designs, but the two sources must remain distinguishable.
For example, a researcher might compare an authentic survey dataset with a synthetic dataset to evaluate the statistical fidelity of a generation method.
That differs from adding artificial records to the observed sample and reporting them as additional participants.
Where datasets are combined, researchers should document the purpose, generation procedure, and analytical treatment of each component.
Record-level provenance indicators may be appropriate when artificial and empirical observations coexist in the same file.
The distinction between synthetic data and real research observations must remain intact throughout data processing and reporting.
How can researchers prevent synthetic data from being mistaken for fabricated or empirical data?
Clear provenance and documentation are essential.
Researchers should identify artificial datasets explicitly, record how they were generated, and explain their intended use.
For projects involving multiple datasets, separate storage or reliable provenance indicators can reduce accidental confusion.
Original empirical records should be preserved according to applicable research requirements.
When synthetic data are deposited in repositories, accompanying documentation should make their artificial origin understandable to downstream users.
Researchers should also ensure that analytical scripts do not unintentionally include synthetic testing records in empirical analyses.
The goal is not merely to avoid accusations of fabrication. It is to preserve an accurate and reproducible account of the evidence supporting the research.
Can AI-generated synthetic data still create research-integrity problems?
Yes. Artificial-data generation can create integrity concerns when the method is misrepresented, the results are overstated, or the generated material is falsely attributed to actual participants or experiments.
For example, AI-generated interview quotations must not be presented as statements made by genuine participants.
Similarly, generated laboratory measurements cannot be reported as instrument readings from experiments that never occurred.
Other problems may involve inadequate validation, misleading statistical interpretation, or failure to disclose consequential AI assistance.
These concerns are not necessarily equivalent to formal misconduct, but they can compromise the reliability of a study.
The broader issue of when AI use becomes a research-integrity problem depends on the honesty, transparency, and methodological validity of the research process.
