03 · What You Need to Know
Practical Safeguards for Keeping AI Simulations Separate From Human Evidence
Why can AI-simulated participants be confused with genuine observations?
AI-generated responses can use the same formats and variables as empirical research data.
For example, a simulated questionnaire dataset may contain participant identifiers, demographic categories, Likert-scale ratings, and open-ended responses.
When imported into statistical software, these records may be indistinguishable from genuine survey data unless their origins are documented.
Similarly, AI-generated interview transcripts may contain realistic quotations, timestamps, and fictional participant identifiers.
The risk increases when files are renamed, merged, copied, or transferred among collaborators.
Researchers should therefore establish safeguards that do not depend exclusively on remembering which records were generated.
What is data provenance?
Data provenance concerns the origin and history of information, including the activities that produced or transformed it.
For AI-simulated participants, provenance may identify the generation model, prompts, hypothetical profiles, creation date, and subsequent processing steps.
For genuine participant data, provenance may include recruitment procedures, collection instruments, and documented data-processing activities.
Preserving these distinctions allows researchers to determine which observations were collected from people and which were generated computationally.
The World Wide Web Consortium's PROV framework provides a general model for describing relationships among data entities, activities, and responsible agents.
Researchers need not necessarily implement that framework formally, but its emphasis on traceable data origins is relevant to research integrity.
What safeguards can prevent confusion?
| Safeguard |
Purpose |
Example implementation |
| Dataset-level labeling |
Identify entire datasets as artificial or empirical |
Descriptive filenames and metadata |
| Record-level provenance |
Distinguish observations when datasets are combined |
A field indicating observed or synthetic origin |
| Separate storage |
Reduce accidental mixing of records |
Distinct directories for simulation and empirical data |
| Generation logs |
Document how artificial records were produced |
Model identifiers, prompts, settings, and creation dates |
| Automated validation |
Detect unexpected artificial records in empirical analyses |
Scripts checking record origin and sample counts |
| Version control |
Track changes to code and documentation |
Recorded revisions and reproducible processing scripts |
| Manuscript disclosure |
Preserve accurate interpretation of findings |
Methods and data availability statements |
These safeguards are complementary. A descriptive filename may help researchers recognize a dataset, while record-level provenance and automated checks can provide stronger protection against accidental mixing.
Should simulated and empirical datasets be stored separately?
Often, particularly when the artificial records are used only for preliminary testing.
Separate storage reduces the likelihood that simulated responses will be included in the final empirical analysis.
For example, a research project might maintain distinct directories for synthetic test data, genuine participant records, and processed analytical datasets.
However, separation alone is not sufficient if files are subsequently copied or merged without preserving provenance.
Researchers should also use descriptive metadata and documented processing procedures.
The appropriate arrangement depends on the complexity of the project and whether combining data sources is methodologically justified.
Should every simulated participant have a provenance label?
Not necessarily when an entire dataset is clearly synthetic and remains separate from empirical records.
However, record-level indicators become especially useful when artificial and genuine observations coexist.
For example, a variable named record_origin may distinguish records labeled observed from those labeled synthetic.
Researchers should define the labels in a data dictionary and preserve them during transformations.
However, a provenance field should not be interpreted as permission to combine artificial and empirical records indiscriminately.
The analytical design must still justify the role of each source.
The principles discussed in labeling synthetic research data apply directly to AI-simulated participant records.
How should simulated participant identifiers be designed?
Identifiers should avoid implying that artificial records correspond to actual recruited individuals.
For example, synthetic records may use a distinct identifier prefix or another documented convention.
However, identifiers alone should not be the only provenance mechanism because they can be modified or removed during processing.
Researchers should also avoid assigning real participant identifiers to generated records.
Doing so may create confusion or incorrectly associate fictional information with identifiable people.
Where simulated profiles are based on genuine participant information, additional confidentiality and data-protection considerations may apply.
Can artificial records be combined with genuine observations?
Potentially, when the research design explicitly justifies the combination and preserves their different origins.
For example, synthetic augmentation may be used during prediction-model development under an appropriate evaluation framework.
However, artificial records should not be counted as additional independently recruited participants.
Researchers should explain how the generated records influence estimation, training, or other analytical procedures.
They should also consider whether combining sources introduces bias, statistical dependence, or information leakage.
The presence of synthetic observations may require analytical treatment different from that used for an entirely empirical sample.
How can automated checks prevent accidental contamination?
Researchers can incorporate validation rules into data-processing scripts.
For example, a script intended to analyze only genuine survey responses may verify that every included record has the expected provenance status.
It may also compare the number of included records with the documented empirical sample.
If artificial records appear unexpectedly, the script can stop rather than silently continue.
Such checks are particularly useful when several collaborators work with shared datasets.
However, automated checks depend on accurate labels and correct implementation.
They should complement rather than replace appropriate data review.
Why should synthetic data be excluded from empirical sample counts?
Because generated records do not represent additional human participation.
For example, a study containing 250 genuine questionnaire responses and 750 AI-generated responses has 250 actual respondents, not 1,000.
Reporting the larger number as a human sample would misrepresent recruitment and the empirical basis of the analysis.
Artificial records may be counted separately as generated observations when relevant to the methodology.
However, their statistical interpretation depends on the generation procedure.
The distinction between synthetic data and real participant data should remain explicit in sample descriptions and analytical reporting.
How should AI-generated interview transcripts be identified?
Simulated interview transcripts should be clearly distinguished from recordings or transcripts obtained from actual participants.
Researchers may identify the files as simulated and document the fictional profiles and prompts used to generate them.
Where excerpts are quoted in publications, their artificial origin should be visible to readers.
For example, a quotation attributed to an AI-simulated teacher should not be presented as testimony from a recruited teacher.
The risk is particularly serious when fictional narratives include realistic contextual details that resemble genuine participant accounts.
Clear attribution preserves the evidentiary distinction between model-generated text and human testimony.
Can AI-generated responses be included in qualitative coding exercises?
Yes, when the purpose is clearly methodological.
For example, fictional interview excerpts may help researchers test code definitions or train coders before analyzing genuine transcripts.
However, these excerpts should remain distinguishable from authentic participant material.
Researchers should avoid including artificial quotations in empirical theme summaries without appropriate identification.
They should also ensure that the coding framework is evaluated against genuine data when the study seeks to interpret actual experiences.
The use of AI-generated responses to test coding frameworks is legitimate only within the evidentiary limits of the exercise.
How should AI simulation be described in the methods section?
The methods should identify the purpose of the simulation and distinguish it from genuine data collection.
Relevant details may include the AI model, available version information, prompts, hypothetical profiles, generation settings, testing period, and number of generated records.
Researchers should explain how artificial responses were stored, labeled, and used.
If genuine human data were also collected, the manuscript should describe those procedures separately.
The reporting should allow readers to determine which findings are based on simulation and which are supported by empirical observations.
A general statement that AI was used may be insufficient when generated records materially affect the analysis.
Should the abstract identify simulated participants?
Generally, when simulation is central to the study design or interpretation of the findings.
For example, an abstract stating that 1,000 students completed a questionnaire would be misleading if all responses were generated by AI.
A more accurate description would identify 1,000 AI-generated respondent records or simulated questionnaire responses.
The abstract should also avoid presenting simulation findings as observed population results.
Researchers should maintain consistent terminology across the title, abstract, methods, results, and conclusions.
Terminological precision is especially important because many readers encounter only the abstract.
Can an AI disclosure statement prevent confusion?
It can contribute to transparency, but it is not sufficient by itself.
For example, a statement that a language model assisted with the research does not establish which observations were generated or how they were analyzed.
Researchers should provide dataset-specific methodological information when artificial records materially affect the study.
AI disclosure and data provenance serve related but distinct functions.
The first identifies relevant AI assistance, while the second explains the origin and history of the observations.
Both may be necessary for accurate interpretation.
How should researchers prevent accidental use of synthetic records in statistical analysis?
Researchers should establish explicit data-selection rules and verify their implementation.
For example, an empirical survey analysis may select only records identified as genuine observations.
The resulting sample size should be checked against recruitment and data collection records.
Researchers may also maintain separate analysis scripts for simulation and empirical datasets.
Where artificial records are intentionally incorporated into a model-development workflow, their role should be documented and evaluated.
The objective is to prevent synthetic observations from entering analyses under assumptions intended only for genuine participant data.
Can synthetic records be accidentally transformed into empirical-looking results?
Yes. Data processing can obscure provenance when source labels are discarded.
For example, a researcher may combine artificial and genuine responses, calculate summary statistics, and export a results table without retaining information about record origin.
The resulting table may appear to describe a single empirical sample.
Researchers should therefore preserve provenance not only in raw datasets but also in analytical outputs and documentation.
Where necessary, separate summaries should be produced for synthetic and empirical sources.
Data lineage should remain traceable from the reported result to the records and procedures that produced it.
What is the role of version control?
Version control can help researchers document changes to code, metadata, and analytical procedures.
For example, a research team may preserve the script used to generate artificial responses and the code that excludes them from empirical analyses.
Changes to these procedures can then be reviewed.
However, version control alone does not guarantee accurate provenance.
Researchers must still establish appropriate naming conventions, metadata, and validation rules.
Confidential participant records should not be placed in public repositories merely to make the workflow reproducible.
Can simulated participants be mistaken for empirical evidence in published datasets?
Yes, particularly when repository documentation is incomplete.
A publicly released dataset containing realistic questionnaire responses may be reused by researchers who do not know that the records were generated.
For example, a dataset titled University Student Engagement Survey may imply genuine data collection if its synthetic origin is not clearly identified.
Repository titles, descriptions, README files, and data dictionaries should therefore communicate the artificial status and intended use.
Researchers should also identify whether the released dataset is the exact material analyzed or an illustrative substitute.
Clear documentation reduces the risk of downstream misuse.
What if a collaborator mistakenly includes synthetic records in an empirical analysis?
The research team should investigate the extent of the error and determine which results were affected.
Relevant datasets, processing scripts, and provenance records should be preserved.
The analysis should be corrected using the appropriate empirical records and documented procedures.
If the error is identified before publication, the manuscript should be revised accordingly.
If affected findings have already been published, appropriate editorial or institutional correction procedures may be necessary.
Accidental inclusion does not automatically establish formal research misconduct, but it can materially undermine the reliability of reported findings.
When does confusion become a research-integrity violation?
Research-integrity concerns arise when the research record inaccurately represents the origin or status of observations.
For example, knowingly reporting AI-generated interview transcripts as genuine participant interviews raises a fabrication concern.
The United States Office of Research Integrity defines fabrication as making up data or results and recording or reporting them.
However, formal findings of misconduct depend on the applicable standards, including evidence of the relevant conduct and culpability.
Researchers should distinguish deliberate misrepresentation from honest error while recognizing that both may require correction.
Strong provenance procedures help prevent confusion and provide evidence for investigating problems when they occur.
How can institutions support proper separation?
Research institutions may establish data-management requirements for synthetic and empirical observations.
These may include standardized metadata, approved storage arrangements, documented generation procedures, and appropriate access controls.
Institutions may also provide guidance on reporting AI-generated records and reviewing analyses that combine artificial and genuine data.
However, governance should remain proportionate to the research risks and methodological context.
A small synthetic dataset used for software testing may not require the same controls as a large AI-generated dataset derived from confidential participant records.
Researchers remain responsible for ensuring that the final research record accurately represents the work conducted.
How should research teams audit provenance before submission?
A pre-submission audit can examine whether reported sample sizes, analytical datasets, and methodological descriptions are consistent.
Researchers may verify that all empirical observations correspond to documented data collection and that artificial records are separately identified.
They should also examine whether summary tables or statistical models inadvertently incorporate synthetic data.
Where AI-generated material is quoted, its source should be accurately attributed.
Finally, the abstract, methods, results, and conclusions should use terminology consistent with the actual data origins.
Such checks can help prevent provenance errors from becoming published scientific claims.
