How to Fact-Check an AI-Assisted Presentation Before You Present It
Short answer: Treat an AI-assisted slide deck as an unverified draft, not as a source. Before presenting it, review every material claim against the original source, recalculate every number, inspect visuals for misleading context, and rehearse the speaker notes as spoken assertions. Keep a simple evidence log so another person can see what was checked, when, and what remains uncertain.
AI can help organize an outline, suggest phrasing, or turn supplied material into a visual sequence. It can also produce plausible but unsupported statements, mismatched citations, distorted charts, and notes that sound more certain than the evidence. The workflow below separates visual polish from factual reliability. It is an educational quality-control method, not legal, tax, financial, privacy, copyright, or compliance advice. For a regulated presentation or a high-stakes claim, consult the relevant qualified professional and the current primary rules.
1. Freeze the draft and define the review boundary
Save the version you intend to review and give it a clear date or version label. Do not fact-check a moving target while an AI tool is still rewriting slides. Then define what the presentation is supposed to establish: its audience, decision or learning objective, time limit, and the claims that would materially change the audience’s understanding.
Mark each slide as one of four types: evidence, interpretation, recommendation, or decoration. Evidence slides need traceable sources. Interpretation slides need a clear distinction between what the source says and what you infer. Recommendations need an owner and an explanation of assumptions. Decorative elements should not imply precision, scale, chronology, or causation.
This boundary matters because a review can miss a claim hidden in a title, caption, chart label, alt text, animation, or speaker note. NIST’s Generative AI Profile frames risk management as an ongoing process of identifying, measuring, and managing risks rather than a one-time approval stamp.[1]
2. Build a claim-and-evidence register
Create one row for every factual statement that a reasonable audience might rely on. Include statements in the slide body, title, subtitle, chart, footnote, image caption, and speaker notes. A useful register has these fields:
- Claim: the shortest testable version of the statement.
- Type: number, date, comparison, quote, product capability, causal explanation, or forecast.
- Source: the exact primary URL, document title, table, page, or section.
- Scope: geography, population, timeframe, unit, and definition.
- Status: verified, revised, unsupported, or intentionally framed as an assumption.
- Owner and date: who checked it and when the check occurred.
Prefer the first-party or original source: a government statistical table for a public statistic, a company’s official documentation for a product feature, an original study for research findings, or a filed report for an organization’s disclosed result. A search snippet, an AI-generated citation, or a secondary summary can help you locate a source, but it should not be the final evidence record.
For each source, verify that the cited passage actually supports the claim as written. A source saying “may” does not support “will.” A source describing a sample does not automatically support a claim about everyone. A source from one year or market does not silently support a different year or market.
3. Verify language before verifying style
Read the slide without its design. Ask whether the wording is narrower or stronger than the evidence. Replace absolute language such as “always,” “proven,” “guaranteed,” or “the best” unless the source and context genuinely justify it. Label forecasts, scenarios, hypotheses, and internal estimates as such. If the claim cannot be tied to a source or a plainly stated assumption, remove it or rewrite it as a question.
Marketing and customer-result slides deserve special care. The Federal Trade Commission says advertising claims must be truthful, not deceptive or unfair, and evidence-based.[2] Its endorsement guidance also emphasizes that endorsements must be honest and not misleading, and that exceptional results should not be presented in a way that implies typical results without adequate support.[3] In practice, do not let AI turn one anecdote, internal experiment, or selected customer comment into a general outcome claim. If the deck is public-facing, have a qualified reviewer assess the current rules for the specific claim and audience.
4. Recalculate every chart and number
Do not accept a chart because its bars look proportionate. Rebuild the calculation from the source data, preferably in a separate spreadsheet or script, and compare the result with the slide. Check totals, percentages, denominators, rounding, unit conversions, date ranges, missing values, and whether categories overlap. Confirm that the axis starts, ends, and scales honestly for the point being made.
Run four tests on every quantitative slide:
- Arithmetic test: Can you reproduce the displayed values from the cited inputs?
- Definition test: Do the source’s terms mean the same thing as the slide’s terms?
- Comparison test: Are the compared groups measured with the same method and timeframe?
- Visual test: Would a reasonable viewer draw a materially different conclusion if the chart were shown as a table?
For public data, record the table or API query, not only the organization’s homepage. The U.S. Census Bureau’s data portal provides access to demographic, economic, and population data along with tools and tutorials; use the underlying table and its definitions when a Census statistic appears in a deck.[4] If a figure is an estimate, label it as an estimate. If a chart is illustrative rather than empirical, say so on the slide.
5. Inspect visuals as evidence, not decoration
AI-generated images can accidentally suggest facts that the text never states. A photorealistic crowd can imply a real event; a map can imply geographic accuracy; a dashboard-like graphic can imply real measurement; and a recognizable person or brand can create an endorsement impression. Ask what a viewer would infer from the image alone.
For each visual, document its purpose and provenance. Is it a conceptual illustration, a photograph, a diagram built from verified data, or an image supplied by a stakeholder? Check labels, legends, icons, arrows, color emphasis, perspective, and cropping. Remove invented UI, fake citations, unreadable pseudo-text, and visual cues that imply a measured result. Use a neutral conceptual illustration when the image is not itself evidence. Also check accessibility: meaningful alt text, sufficient contrast, readable type, and a spoken explanation for important visual information.
6. Review structure and speaker notes separately
Once claims and visuals are checked, review the narrative. Every transition should answer why the next slide follows. Distinguish observation from interpretation and interpretation from recommendation. Look for duplicated claims, missing caveats, unsupported causal arrows, and a conclusion that is stronger than the body of evidence.
Then read the speaker notes aloud without looking at the slide. Notes often contain the most consequential unsupported language because they are written as conversational shortcuts. Highlight predictions, superlatives, customer stories, product capabilities, and phrases such as “this shows” or “therefore.” Attach a source or an explicit qualification to each one. If the presenter must say that something is unknown, make that uncertainty easy to say naturally rather than hiding it in a footnote.
7. Use a bounded rehearsal test
Conduct a short rehearsal with a reviewer who did not create the deck. Give that person the audience brief and ask them to flag three things: a claim they would want to verify, a chart they interpret differently from the intended message, and a caveat they think is missing. Do not prompt them with the answer. Record the slide number and revise the deck, the notes, or the evidence register.
This is not a guarantee that the presentation is correct. It is a practical way to test whether the audience can distinguish evidence, interpretation, and uncertainty. Rehearse again after material edits, because a shortened sentence or redesigned chart can change the meaning.
8. Original go/no-go checklist
Use this decision tool immediately before delivery. Mark each item yes, no, or not applicable.
- Every material factual claim has an exact, accessible source.
- The source passage matches the claim’s wording, scope, date, and certainty.
- Every number, percentage, total, and conversion has been independently recalculated.
- Charts use honest labels, units, axes, and category definitions.
- Illustrations do not imply real people, events, measurements, endorsements, or fabricated interfaces.
- Speaker notes contain no stronger claim than the slide or source supports.
- Assumptions, estimates, forecasts, and open questions are visibly labeled.
- A second reviewer has completed a bounded rehearsal or read-through.
- Unresolved items are removed, narrowed, or explicitly assigned for follow-up.
Decision rule: proceed only when every applicable item is “yes.” If any item is “no,” place the deck on hold or revise the affected slide. “Not applicable” should be explained in the evidence register, not used to avoid review. Keep the register with the final deck so the next update starts from traceable evidence rather than from memory.
Material caveats
Sources can change, links can move, datasets can be revised, and a correct statement can become misleading when its timeframe or audience changes. A fact-check is therefore a review of a particular version for a particular use, not a permanent certification. Current advertising, accessibility, data-use, intellectual-property, or sector-specific requirements may also depend on jurisdiction and context; obtain qualified advice when those issues matter to the presentation.
