03 · What You Need to Know
What Researchers Should Verify Before Delivering an AI-Assisted Presentation
Why Is Presentation Verification a Research Integrity Issue?
Research presentations communicate scientific findings to audiences who may rely on them for subsequent research, teaching, professional practice, or policy decisions.
Although a conference presentation is often shorter than a journal article, it remains a public representation of research evidence.
Errors in slides can mislead audiences even when the underlying manuscript is accurate.
Generative AI introduces additional opportunities for such errors because it may summarize information incorrectly, invent supporting details, or alter the meaning of figures during redesign.
Researchers should therefore evaluate the presentation as a scientific communication product rather than merely a visual document.
What Should Be the Authoritative Source for Verification?
The appropriate reference depends on the claim being checked.
For methodological statements, the study protocol, manuscript, and research records may provide the relevant evidence.
For statistical findings, verified analytical outputs and results tables are generally more authoritative than AI-generated summaries.
For scientific figures, the underlying data and plotting code may be necessary.
For external claims, the original cited publications should be consulted.
The objective is to establish a traceable relationship between what the presentation communicates and the evidence supporting it.
What Are the Main Areas of Presentation Verification?
| Area |
Potential AI-related problem |
Verification source |
| Research objectives |
Changed or invented research questions |
Approved manuscript or protocol |
| Methodology |
Incorrect sampling, design, or analytical procedures |
Methods section and research records |
| Statistical results |
Incorrect values or interpretations |
Verified analytical outputs |
| Figures and diagrams |
Altered data, scales, or relationships |
Source data, plotting code, or approved framework |
| References |
Fabricated or misattributed citations |
Original publications |
| Conclusions |
Exaggerated certainty or causality |
Results, design, and discussion |
| Speaker notes |
Unsupported oral explanations |
Verified manuscript and presentation |
| Disclosure |
Missing or inaccurate AI acknowledgment |
Applicable institutional and venue policies |
These checks should be adapted to the research design and presentation context. Not every presentation contains statistical analyses, and some require additional verification of qualitative findings or technical demonstrations.
How Should Researchers Verify Slide Claims?
Each consequential scientific claim should be traceable to the manuscript, research records, or an appropriate external source.
For example, a slide stating that an intervention improved student performance should be checked against the actual study design and findings.
If the study reported an association rather than a causal effect, the slide should preserve that distinction.
Researchers should examine not only body text but also slide titles, annotations, figure captions, and concluding statements.
A misleading headline can change the audience's interpretation even when the figure beneath it is numerically correct.
Why Must Statistical Results Be Checked Independently?
Generative AI may reproduce numerical information incorrectly or misunderstand its statistical meaning.
For example, a coefficient of 0.08 may be reported as 0.8, or a confidence interval may be interpreted as containing a specified proportion of individual observations.
AI may also describe nonsignificant findings as proof that no effect exists.
Researchers should compare reported values directly with verified statistical outputs.
Effect measures, reference categories, uncertainty intervals, and model specifications should be checked where relevant.
The American Statistical Association's guidance on p-values is particularly relevant when reviewing claims that equate statistical significance with scientific importance.
How Can Researchers Detect Methodological Misrepresentation?
AI may simplify methodological descriptions in ways that change the study's evidentiary status.
For example, a quasi-experimental design may be described as a randomized experiment, or a convenience sample may be presented as representative of a national population.
Such changes can affect interpretations of internal validity, generalizability, and causality.
Researchers should verify descriptions of participant recruitment, allocation, measurement, data collection, and analysis.
When methods are condensed for presentation, the remaining explanation should preserve distinctions necessary to evaluate the findings.
Should Researchers Verify Which Findings Were Selected?
Yes. Verification should examine not only whether the included results are accurate but also whether the selection creates a misleading impression.
AI may prioritize statistically significant or visually dramatic findings while omitting inconclusive primary outcomes.
For example, a presentation may emphasize improvements in student satisfaction while failing to mention that the primary learning outcome showed little evidence of improvement.
The selected findings may be individually correct, yet their relative emphasis can distort the study's overall conclusions.
Researchers should compare the presentation with the complete results and apply the principles of selecting important research findings responsibly.
How Should Scientific Graphs Be Verified?
Data-bearing graphs should be checked against the underlying dataset or verified analytical outputs.
Researchers should examine plotted values, sample sizes, statistical summaries, uncertainty intervals, axes, labels, and graphical encodings.
For example, a bar chart may use correct numerical values while employing a truncated baseline that exaggerates apparent differences.
Similarly, an AI-redesigned coefficient plot may contain altered confidence intervals or incorrect legend mappings.
The process of verifying AI-modified research visualizations should be completed before figures are incorporated into the final presentation.
How Should Scientific Diagrams Be Verified?
Scientific diagrams may communicate theoretical relationships, methodological procedures, or technical systems rather than numerical data.
AI can introduce incorrect components, reversed arrows, or unsupported connections.
Researchers should compare these diagrams with their authoritative source materials.
For a conceptual framework, the relevant reference may be the approved theoretical model. For a methodological flowchart, it may be the study protocol.
Researchers should verify every consequential label, relationship, and sequence.
The possibility of AI-generated scientific diagram inaccuracies makes this check particularly important when diagrams are used to explain mechanisms or research procedures.
Why Should References Be Verified?
Generative AI can produce citations that appear academically plausible but do not correspond to actual publications.
It may also cite real sources that do not support the associated claims.
Researchers should verify that each cited work exists, that its bibliographic details are accurate, and that it supports the statement being made.
Where a slide uses a figure or quotation from another publication, attribution and licensing requirements should also be considered.
References should not be accepted merely because they contain familiar author names, journal titles, or digital object identifiers.
Can AI-Generated Slide Titles Be Misleading?
Yes. Slide titles often function as concise statements of the main message.
AI may strengthen these statements to make the presentation more engaging.
For example, "Online Learning Improves Academic Performance" communicates a causal claim, while "Online Learning Engagement Was Associated With Academic Performance" describes an observational relationship.
The difference may be only a few words, but it changes the scientific interpretation.
Researchers should verify headlines as carefully as statistical statements and figure captions.
Should Speaker Notes Be Included in Verification?
Yes. Speaker notes may contain scientific explanations that are not visible on the slides.
AI may invent mechanisms, add unsupported examples, or misinterpret statistical results while drafting oral narration.
For example, a note may attribute an observed association to increased motivation even though motivation was not measured.
Researchers should verify AI-generated research presentation speaker notes against the original study.
Oral statements should meet the same standards of scientific accuracy as visible slide content.
How Can Researchers Check for Oversimplification?
AI-generated slides often contain less text than the source manuscript, which can improve readability.
However, excessive compression may remove information essential to interpretation.
Researchers should check whether the presentation preserves the study design, relevant statistical uncertainty, and limitations affecting generalizability or causal inference.
The issue is not whether every methodological detail appears on screen.
It is whether audiences receive enough information to understand what the findings establish and what remains uncertain.
The risk of AI oversimplifying research presentations should be evaluated alongside factual accuracy.
Can AI Verify Its Own Presentation?
AI can help identify inconsistencies, review slide text, and compare claims with supplied source materials.
However, an AI-generated assurance that a presentation is accurate should not be treated as independent verification.
The same model may overlook errors it introduced or provide a confident assessment without checking the underlying evidence.
Researchers should establish verification criteria independently of the generated presentation.
For consequential numerical claims, direct comparison with analytical outputs is more dependable than model reassurance.
Should Researchers Use a Claim-to-Source Verification Matrix?
A claim-to-source matrix can make verification more systematic.
Researchers can record each consequential claim, its slide location, supporting source, and verification status.
For example, a statement about an intervention effect can be linked to the relevant statistical table and manuscript paragraph.
This approach is particularly useful for presentations containing many numerical findings or claims from external literature.
AI may assist with preparing the matrix, but the researcher should confirm that the cited sources genuinely support the claims.
The matrix need not appear in the presentation itself. It can remain part of the preparation and quality-control record.
How Should Qualitative Research Presentations Be Verified?
Qualitative findings require attention to the relationship between themes, participant accounts, and analytical interpretation.
AI may simplify themes excessively, remove contextual variation, or invent quotations that resemble plausible participant statements.
Researchers should verify that presented themes correspond to the actual analysis and that quotations are reproduced accurately.
Where participant confidentiality is relevant, quotations and contextual details should be reviewed for disclosure risks.
AI should not transform illustrative participant accounts into claims about the entire study population without analytical justification.
What About Mixed-Methods Presentations?
Mixed-methods studies require verification of both quantitative and qualitative findings and the relationship between them.
AI may overemphasize one strand of evidence or introduce a causal explanation that the integrated analysis does not support.
Researchers should confirm that the presentation preserves the intended integration, including areas of convergence, divergence, or complementarity.
Figures combining different forms of evidence should be checked for accurate labels and interpretations.
How Should Accessibility Be Verified?
Scientific accuracy is essential, but audiences must also be able to access the information.
Researchers should inspect text size, contrast, reading order, and whether essential distinctions depend exclusively on color.
Complex figures may require verbal explanations or accessible descriptions.
When presentations are distributed digitally, the document structure and alternative text should be reviewed according to the capabilities of the file format.
AI can assist with accessibility improvements, but automated suggestions do not replace practical evaluation.
Why Must the Final Rendered Slides Be Inspected?
Presentation files can change appearance when opened in different software environments or exported to PDF.
Fonts may be substituted, mathematical symbols may render incorrectly, and text or figures may become clipped.
Animations, transitions, and embedded media may also behave differently across platforms.
Researchers should inspect the version that will actually be displayed.
A scientifically correct editing file is not sufficient if the projected or exported version omits important information.
Should AI Assistance Be Disclosed?
Disclosure depends on the relevant conference, institutional, publisher, and professional requirements.
Researchers should identify whether AI was used for language editing, substantive content generation, figure creation, or other consequential activities.
The question of disclosing AI-generated research presentation graphics may require separate consideration when generated images or diagrams are included.
Disclosure should accurately describe the nature of AI involvement rather than rely on a vague acknowledgment.
Regardless of disclosure requirements, researchers remain responsible for the accuracy of the presentation.
What Should Researchers Do When Verification Reveals an Error?
The response should depend on the nature and consequences of the discrepancy.
A typographical error may require a straightforward correction.
An incorrect statistical estimate should be checked against the analytical output and corrected wherever it appears, including captions and speaker notes.
A misleading causal conclusion may require revising the presentation's central argument.
Researchers should not simply ask AI to make the questionable statement sound more cautious without determining what the evidence actually supports.
When a discrepancy cannot be resolved before presentation, the unsupported claim should be removed or clearly qualified rather than presented as verified.
