03 · What You Need to Know
How to Use Generative AI Without Misrepresenting Research Findings
Why is the Results section particularly sensitive to AI-generated errors?
The Results section reports what the research produced. In quantitative studies, this may include descriptive statistics, estimated relationships, hypothesis tests, confidence intervals, and model outputs. In qualitative studies, it may present themes, categories, patterns, participant accounts, or interpretations developed through the analytical process.
These findings are not interchangeable with plausible statements about what similar studies might discover. They must be grounded in the specific evidence collected and analyzed.
Generative AI can produce fluent descriptions of results even when the underlying information is incomplete. It may fill missing details with plausible numbers or expected patterns, particularly when asked to write a complete section from a research title or hypothesis.
A scientifically accurate Results section must report the findings actually obtained, including results that contradict the researcher's expectations.
What can AI appropriately do when drafting research results?
AI assistance is most useful when the findings have already been established through a documented analytical process. The model can then support communication rather than independently determine what the research found.
| Results-writing task |
Potential AI assistance |
Required verification |
| Organizing findings |
Suggesting a logical sequence based on research questions or outcomes |
Completeness and appropriate ordering of reported findings |
| Descriptive statistics |
Converting verified tables into readable sentences |
Means, frequencies, percentages, dispersion, and sample sizes |
| Inferential statistics |
Improving descriptions of documented tests and model outputs |
Test statistics, estimates, intervals, p-values, and directions of effects |
| Qualitative findings |
Improving descriptions of researcher-established themes |
Theme definitions, evidence, participant quotations, and contextual meaning |
| Tables and figures |
Suggesting captions or explaining verified visualizations |
Consistency with the underlying data and graphical content |
| Language revision |
Improving clarity, concision, and transitions |
Preservation of numerical and interpretive meaning |
These applications should not be confused with asking AI to conduct the analysis itself. AI-assisted analysis may be possible in particular workflows, but it introduces separate requirements involving methodological validity, reproducibility, data governance, and disclosure.
The broader distinction between AI assistance across manuscript sections is useful here because the Results section has an especially direct obligation to represent empirical evidence.
Why can AI misreport statistical results?
Statistical outputs contain relationships between numbers, labels, assumptions, and analytical procedures. A model may recognize the appearance of a regression table without correctly identifying the dependent variable, reference category, confidence interval, or meaning of a coefficient.
Errors can include confusing standard deviations with standard errors, interpreting an odds ratio as a probability, reversing the direction of a relationship, or reporting statistical significance where the analysis does not support it.
Consider a regression coefficient of 0.24. Its interpretation depends on the model specification, measurement scales, coding, and whether it is standardized. The number alone does not establish a 24% increase in an outcome.
Similarly, a p-value does not measure the probability that the null hypothesis is true. AI-generated descriptions that treat it that way require correction.
Researchers should compare each statistical statement with the original output and the analytical procedure that produced it. The relevant question is not simply whether the number was copied correctly, but whether its meaning was represented accurately.
Can AI write results from a statistical table?
Yes, provided the table is verified and contains enough information to support the requested description. A model may help convert tabular results into prose that emphasizes the principal findings without mechanically repeating every cell.
However, a table may omit information needed for interpretation. For example, a coefficient table might not identify the coding of categorical predictors or the model's adjustment variables. If those details are missing, the generated prose may be misleading even when the numerical values are reproduced correctly.
A useful approach is to provide the verified table together with relevant definitions and analytical context. The researcher should instruct AI not to infer missing statistics or describe relationships beyond those supported by the supplied information.
The resulting paragraph must still be checked against the original analysis. A prompt that requests accuracy cannot guarantee it.
What about null, nonsignificant, or unexpected results?
Generative AI may produce narratives that emphasize positive or statistically significant findings because such narratives resemble familiar research reporting patterns. This can be problematic when the study produced mixed, null, or unexpected results.
A nonsignificant finding should not automatically be described as evidence that no relationship exists. The interpretation depends on the estimate, uncertainty, study design, sample size, and other relevant considerations.
Likewise, a statistically significant result should not automatically be characterized as practically important. Statistical significance and substantive importance address different questions.
Researchers should ensure that AI-assisted drafting preserves the full pattern of findings, including outcomes that do not support the original hypotheses. Selective reporting can mislead readers even when the individual statements retained in the manuscript are numerically correct.
Can AI draft qualitative research findings?
AI may help organize or refine descriptions of themes already developed through a documented qualitative analysis. It may also assist with formatting thematic summaries or improving transitions between findings.
However, qualitative results depend on contextual interpretation. Participant statements cannot necessarily be reduced to their most common wording, and a theme is not automatically established because an AI system identifies recurring expressions.
Researchers must verify that themes correspond to the analytical process, that quotations are authentic, and that the relationship between participant accounts and interpretive claims remains defensible.
AI should not invent quotations, combine statements from different participants into a fabricated quotation, or present synthetic participant accounts as authentic evidence.
Confidentiality is also important. Interview transcripts and other qualitative materials may contain identifying or sensitive information. Researchers should not upload protected data to an AI service without confirming that the relevant ethical approvals, consent arrangements, institutional requirements, and data-handling safeguards permit it.
Where does reporting end and interpretation begin?
The boundary between Results and Discussion varies by discipline and article format. Some journals expect a relatively descriptive Results section, while qualitative and mixed-methods publications may integrate findings and interpretation more closely.
Nevertheless, researchers should distinguish statements directly supported by the analysis from explanations requiring additional reasoning.
Reporting a finding
The analysis identified a positive association between platform participation and student engagement.
Explaining a finding
The association may reflect differences in motivation, instructional support, or other factors that require further examination.
The second statement proposes possible explanations. These may be appropriate for consideration, but the association alone does not establish them.
When using AI to develop the Discussion section, the researcher must evaluate such explanations against the evidence and relevant literature.
Can AI generate or modify figures for the Results section?
AI may assist with writing plotting code, selecting appropriate visualization formats, or preparing captions. However, figures that represent research data should be derived from verified numerical or image evidence through an inspectable process.
An image-generation system may produce a convincing graph without correctly plotting the underlying observations. It may invent error bars, alter proportions, or add unsupported significance annotations.
Researchers should therefore distinguish AI-assisted plotting from synthetic chart imagery. Where a figure has been generated without a verifiable connection to the research data, regenerating the publication figure from verified data is the more defensible approach.
Journal policies on AI-generated visual material may also impose additional restrictions.
What if AI generates results that appear reasonable but were not supplied?
Unsupported results must not be included as empirical findings. A model may invent a mean difference, correlation coefficient, thematic category, or statistical conclusion that fits the study's topic but does not originate from the actual analysis.
The fact that a generated result appears plausible does not make it an observation. Nor does later editing convert fabricated content into valid research evidence.
Watch Out
Never ask generative AI to supply missing results, produce favorable statistical findings, or create participant quotations for presentation as actual research evidence. If the analysis is incomplete, complete or appropriately revise the analysis rather than filling the manuscript with generated findings.
Do journals permit AI-assisted Results section drafting?
Publisher policies differ in their treatment of language editing, substantive manuscript drafting, and AI used in research analysis. Authors should check the current policy of the intended journal before using AI-generated scientific prose.
ICMJE emphasizes that human authors remain responsible for AI-assisted content and should disclose relevant use according to its recommendations. COPE likewise maintains that AI tools cannot assume authorship responsibilities.
Where AI contributes to statistical analysis, qualitative coding, or another research procedure, its role may need to be described in the Methods section with sufficient information to understand and evaluate the process.
AI assistance with writing the Results section does not remove the requirement to verify the analysis or report findings accurately.
