03 · What You Need to Know
How Generative AI Can Facilitate Research Data Fabrication
What is research data fabrication?
Research data fabrication involves creating data or results that were not genuinely obtained and recording or reporting them as research evidence.
The United States Office of Research Integrity (ORI) defines fabrication as making up data or results and recording or reporting them.
This definition focuses on the false representation of research evidence. It does not depend on whether the invented information was produced manually, through a spreadsheet, by statistical software, or using generative AI.
For example, a researcher who invents laboratory measurements and enters them into a dataset may be fabricating data. A researcher who instructs AI to produce those measurements and then records them as actual laboratory results engages in the same underlying type of misrepresentation.
Under formal misconduct frameworks, findings also require the relevant evidentiary and culpability standards to be satisfied. The distinction between a suspected fabrication and an established misconduct finding should be preserved.
What kinds of research data can generative AI fabricate?
Generative AI can produce many forms of material resembling research observations.
The output may be textual, numerical, visual, or structured, depending on the system and the instructions supplied.
| Research material |
What AI can generate |
Integrity concern if misrepresented |
| Survey data |
Responses, demographic variables, rating scales, and open-ended answers |
Invented responses may be presented as data collected from actual respondents. |
| Interview transcripts |
Dialogue, participant narratives, and quotations |
Generated statements may be falsely attributed to real interview participants. |
| Laboratory measurements |
Numerical readings, observations, and experimental records |
Values may be reported as measurements from experiments that were never conducted. |
| Scientific images |
Images resembling microscopy, medical imaging, or experimental observations |
Generated visuals may be presented as authentic evidence. |
| Statistical results |
Tables, coefficients, descriptive statistics, and analytical summaries |
Invented outputs may be reported as results of analyses that were not performed. |
| Qualitative field notes |
Descriptions of observations, events, and participant behavior |
Fictional accounts may be represented as actual field observations. |
The ability to produce these materials does not mean every generated example will be convincing or internally consistent. AI outputs may contain contradictions, implausible distributions, or other errors.
Nevertheless, superficial plausibility is not sufficient to establish authenticity.
Why can AI-generated fabricated data appear credible?
Generative AI systems can reproduce linguistic and statistical patterns learned during training or represented in supplied examples.
For textual data, this may produce interview responses that resemble ordinary human speech, including variation in vocabulary, sentence length, and expressed opinions.
For numerical data, systems or AI-assisted workflows may generate values within plausible ranges or according to specified distributions.
These characteristics can make artificial observations appear consistent with a research context.
However, plausibility and authenticity are different properties.
A realistic-looking interview transcript does not establish that an interview occurred. A numerical dataset with plausible means and standard deviations does not establish that measurements were collected.
The research record must support the origin of the observations independently of their appearance.
Does AI make research fabrication easier than manual fabrication?
Generative AI may reduce the time and technical effort needed to produce large quantities of plausible-looking material.
For example, manually inventing numerous open-ended survey responses can be laborious. A generative system may produce many varied responses rapidly.
AI may also assist in formatting invented observations into tables, datasets, or transcripts resembling legitimate research records.
However, the extent to which AI makes fabricated research harder to detect is an empirical question rather than an established universal conclusion.
Detection depends on the type of data, available provenance records, independent verification, and the methods used to assess authenticity.
The distinction between AI-assisted and manual research data fabrication concerns the tools and mechanisms involved, not a different ethical definition of fabrication.
When do AI-generated survey responses become fabricated research data?
AI-generated survey responses become a fabrication concern when they are recorded or reported as though they were supplied by actual participants.
Suppose a study recruits 150 participants but the researcher generates 50 additional responses through AI and combines them with the authentic dataset.
If the resulting manuscript reports 200 human participants without distinguishing the generated observations, the research record misrepresents the sample.
This affects more than the sample size. The reported descriptive statistics, relationships among variables, and conclusions may depend partly on invented observations.
The same concern applies when researchers use AI to generate missing participant records without a legitimate and disclosed methodological basis.
Appropriate statistical imputation is not equivalent to secretly inventing additional participants. Imputation estimates missing values under specified assumptions and should be reported accordingly; it does not establish that new participants were observed.
The specific issue of AI-generated fake survey responses warrants particular attention because structured questionnaires can be reproduced in a superficially credible format.
Can AI-generated interview transcripts constitute fabrication?
Yes, when fictional dialogue is presented as a transcript of an interview that did not occur.
Generative AI can simulate how a participant might respond to questions about education, healthcare, workplace experiences, or other subjects.
Such material may be useful for pilot testing interview protocols or developing coding exercises, provided that it is clearly identified as simulated.
However, reporting an AI-generated quotation as the statement of a genuine participant falsely attributes evidence to a person or interview event.
This is particularly consequential in qualitative research because the meaning of a quotation depends partly on its relationship to the participant, research context, and actual interaction.
Researchers should preserve appropriate source records and distinguish AI-generated participants or interview transcripts from authentic qualitative evidence.
Can AI fabricate laboratory or experimental results?
Generative AI can produce numerical values and technical descriptions resembling laboratory results, but generated numbers are not measurements merely because they use appropriate scientific units or fall within plausible ranges.
A table containing temperature readings, concentrations, reaction times, or instrument outputs must be traceable to actual experimental procedures if it is presented as empirical evidence.
The absence of corresponding experiments, instrument records, or other supporting documentation may raise serious questions about authenticity.
However, missing documentation alone does not conclusively prove fabrication. Records can be lost, incomplete, or maintained in different systems.
The relevant assessment requires examining the available evidence and the researcher's explanation.
Researchers should distinguish generated examples used for instrument testing or simulation from measurements claimed to have been observed.
What is the difference between fabricated data and legitimate synthetic data?
Synthetic data are artificially generated observations created for a defined purpose. Depending on the method, they may reproduce selected statistical properties of real data or represent simulated scenarios.
Such data can support legitimate activities including software testing, methodological development, privacy-related experimentation, and simulation studies.
Fabrication, by contrast, concerns falsely recording or reporting invented data or results as genuine research evidence.
Legitimate synthetic data
Artificially generated material whose origin, purpose, and methodological limitations are accurately identified and reported.
Fabricated empirical data
Invented observations or results falsely recorded or reported as though they came from actual participants, experiments, measurements, or other research activities.
The same artificial dataset could be appropriate in a software-testing exercise and unacceptable if presented as genuine participant responses.
Its ethical status depends on how it is created, used, documented, and represented.
The distinction is explored further through legitimate AI-generated synthetic research data and the difference between synthetic and fabricated data.
Can fabricated data be mixed with real observations?
Yes, and doing so can make the resulting dataset particularly misleading.
When artificial records are added to genuine observations without accurate labeling or methodological justification, the final dataset no longer represents the reported empirical sample.
The problem is not resolved simply because most observations are authentic.
For example, adding 20 invented records to a dataset containing 480 genuine participants would still misrepresent the total number of actual participants if all 500 were reported as observed.
Researchers may legitimately combine empirical and synthetic information in certain methodological designs, but the provenance and role of each component must be disclosed and scientifically justified.
The analysis should not imply that simulated observations provide additional independent empirical participants.
Can AI-generated fabricated data distort statistical conclusions?
Yes. Invented observations can change estimated means, variances, correlations, regression coefficients, confidence intervals, and other analytical results.
The direction and magnitude of the distortion depend on the fabricated values and the analytical method.
Even when fabricated records resemble the distribution of genuine data, they do not represent additional independent observations from the target population.
Consequently, treating them as authentic can produce misleading sample sizes, precision estimates, and inferential claims.
The problem may also affect qualitative conclusions when generated statements introduce themes or experiences that were never reported by participants.
Research findings must remain traceable to the evidence that actually supports them.
Can researchers detect AI-fabricated datasets reliably?
There is no universal method that can reliably determine whether every dataset was fabricated using AI.
Statistical irregularities, duplicated patterns, implausible distributions, or inconsistent metadata may provide reasons for further investigation.
However, the absence of such indicators does not establish authenticity, and their presence does not independently prove misconduct.
Provenance evidence is often more informative than attempting to infer fabrication from the final values alone.
Depending on the study, useful records may include original survey platform exports, recruitment logs, instrument files, laboratory notebooks, data collection timestamps, and documented analytical transformations.
These records should be evaluated proportionately and with appropriate privacy safeguards.
The limitations of detecting AI-fabricated research make preventive documentation especially important.
Who is responsible when AI is used to fabricate research data?
Human researchers remain responsible for the accuracy of the research evidence they record and report.
An AI system does not become an accountable research participant or author merely because it generated the false observations.
The International Committee of Medical Journal Editors emphasizes that human authors are responsible for reviewing and ensuring the accuracy and integrity of material produced with AI assistance.
However, individual responsibility in a formal misconduct proceeding must be established under the applicable procedures and evidentiary requirements.
Researchers should distinguish general responsibility for the research record from a formal determination that a particular individual intentionally, knowingly, or recklessly engaged in misconduct.
The principle of human accountability when AI performs a prohibited research action remains central.
