Plan the reading order
Supply the audience, decision, source material and required coverage. Review the outline before generating. For a QBR, keep results, variance, risks and next decisions distinct instead of mixing them into a status collage.
AI deck citation guide
How to Cite Sources in an AI-Generated Presentation
Citing an AI-generated presentation is not just a bibliography task. Teams need to disclose AI assistance, verify factual claims against real sources, attribute generated visuals and data, and make the human contribution clear enough that the deck can be trusted.
Lumen · real output
Read the deck before you choose the workflow.
Open an actual exported deck, then compare its outline, source constraints and slide choices with the job you need to finish.
Governing Cybersecurity Risk: An Executive Briefing on NIST CSF 2.0
A 10-slide briefing for an organisation’s leaders on NIST’s Cybersecurity Framework 2.0, built only from the framework document: the six Functions as a set with Govern at the centre, what Govern covers, questions leaders can ask their teams, how Current and Target Profiles show the gap to close, and the four Tiers. It describes no company, so it has no performance data.
Inspect the source brief or data
A briefing for the executive leadership of an organisation of any kind; no company is named and none is described. Cover, in this order: what the framework is and why leaders should care; the six Functions, shown together as a set with Govern at the centre, with no numbering or order because the document says they are addressed concurrently; what the Govern Function covers, using the document's own list of Govern topics; questions leaders can ask their teams, in the document's own terms; what Current and Target Profiles are and how comparing them shows the gap to close; the four Tiers by name; and what to ask for next. Do not show any process as numbered steps. Say executives, leaders or the organisation wherever the document does: it mentions boards of directors only as readers, so do not give a board any role the document does not give it. Give at most one slide to how different kinds of organisation use the framework. Use only the supplied document. Where it gives no figure, date or finding, do not supply one: use no placeholders, no illustrative numbers and no made-up status data. Pictures should show technology, infrastructure or abstract subjects, not people. Workflow: Corporate executive. 10 slides requested. Document supplied as text (NIST, "The NIST Cybersecurity Framework (CSF) 2.0, NIST CSWP 29 (26 February 2024)", https://nvlpubs.nist.gov/nistpubs/CSWP/NIST.CSWP.29.pdf).
Read the sample review notes
A separate reviewer agent read the deck’s 70 claims, slides and speaker notes, against the framework document: 62 are supported, 8 loosen or go a little beyond its wording, and none is missing from it or contradicts it. 19 of the 70 are advice the deck gives, and all 19 are supported. It is not a person’s review.
Refine at the right scope
Edit structured slide fields, preview field alternatives or review a whole-slide rewrite. Ask Lumen can propose wider changes; inspect and select the changes before applying them to the saved deck.
Inspect the actual handoff
Check charts, speaker notes, images and contrast in presentation view and the downloaded file. A new-period deck is a separate draft with changed values and sources to review.
Outline planning and field alternatives
Choose a creation mode and workflow profile, supply source material and review the planned coverage. Slide fields have layout-specific budgets. Field assistance offers alternatives with rationale and word counts; whole-slide refinement shows field-level changes. Ask Lumen can propose wider structural edits, which require a reviewed preview before application.
Data, themes and the next reporting period
Paste spreadsheet rows, choose header handling and preview replacement chart data. Where offered, select Business metrics explicitly and inspect their source receipt. Review changes in values and sources before accepting a next-period draft. Theme controls cover brand colors, type, logo placement, background and contrast choices; a custom theme still needs visual review.
Presenting and the exported deck
Inspect full-screen and audience views, speaker notes, next-slide preview, navigation and timer. Check PDF and PowerPoint separately. Supported elements can become native editable shapes or charts; complex visuals can use a flattened fallback. Speaker notes and per-slide handoff choices should be inspected in the actual file.
PowerPoint export is hybrid: supported content may remain editable while complex visuals can be flattened. Conversion between a responsive story and slides requires a preview; it is not lossless.
The four citation duties
An AI deck can involve human writing, model assistance, source material, generated images, and derived data. Each layer needs a different treatment.
1. Cite the AI model as a tool
If AI materially helped draft, summarize, visualize, or restructure the deck, disclose that assistance in a speaker note, appendix, methodology slide, or internal review note.
2. Describe the human contribution
Make clear who supplied the prompt, chose the sources, edited the output, validated the claims, and approved the final message. Authorship and accountability remain human.
3. Verify and cite real sources
Do not cite the AI system as the source of a market size, financial figure, quote, regulation, or case study. Verify the claim and cite the underlying primary or trusted secondary source.
4. Attribute visuals and data
Generated images, recreated charts, adapted tables, and transformed data still need attribution rules. Teams should track what was generated, what was sourced, and what was edited.
5. Preserve reproducibility
For high-stakes decks, keep prompts, source files, review notes, and assumptions available so another reviewer can understand how the slide was produced.
6. Match the audience standard
An internal brainstorm may need light disclosure. A board deck, investor deck, compliance deck, or client deliverable usually needs explicit source references and review ownership.
A practical AI contribution and citation framework
1
Label the AI role
Was AI used for ideation, outline generation, source summarization, slide drafting, image generation, chart explanation, or editing? Name the role so reviewers know where to look.
2
Separate facts from wording
AI can help word a claim, but the claim still needs a source. Keep a source list for metrics, quotes, external facts, research findings, and legal or regulatory statements.
3
Attach citations close to the claim
Use footnotes, speaker notes, appendix references, or slide-level source fields. The point is not academic ornamentation; it is making review fast and defensible.
4
Record final human approval
Before presenting, assign an owner for factual accuracy and final messaging. AI involvement should not blur who is accountable for the deck.
What to cite, and what not to cite
| Deck element | What belongs in the citation trail | Common mistake |
|---|---|---|
| Market data | The original report, database, filing, analyst note, or approved internal estimate. | Citing the AI output instead of the source that supports the number. |
| Customer quotes | The transcript, interview note, CRM record, or approved testimonial source. | Letting AI paraphrase the quote without preserving meaning or consent boundaries. |
| Generated visuals | The fact that the visual was AI-generated, plus the tool and prompt context if policy requires it. | Treating generated visuals as source evidence rather than illustrative material. |
| AI-written summaries | The source documents summarized and the human reviewer who checked the synthesis. | Assuming polished language means verified analysis. |
Citing a deck you built with ChatGPT, Claude, or Gemini
The most common version of this question is practical rather than academic: someone asked a general-purpose chat model to write a PPT, pasted the result into PowerPoint or Google Slides, and now has to present it. The citation problem in that workflow is not the model — it is that the sources never travelled with the text. By the time the outline is slides, nothing on the page records where a market figure or a customer statistic came from.
Handle it in this order. First, separate what the model wrote from what the model asserted. Wording is a drafting aid and needs a disclosure, not a citation. Assertions — a market size, a growth rate, a regulation, a quote — need a real source found and checked independently, because a chat transcript is not evidence for a business fact.
Second, disclose the assistance where your audience's standard requires it. An internal working session may need nothing; a board deck, an investor deck, a compliance deck, or a graded submission usually needs an explicit note naming what the AI did — ideation, summarisation, drafting, image generation, chart explanation — and who reviewed the result.
Third, attach the citation to the claim rather than to the deck. A single “made with AI” line in the appendix tells a reader nothing about whether slide 14's number is real. Footnote the slide, or keep a slide-level source field, so that the check is one glance rather than an archaeology exercise.
The workflow lesson is that citation is cheap while you are still holding the sources and expensive once you are holding only slides. That is the case for keeping generation, source material, and review in the same place rather than moving text between a chat window, a document, and a deck.
Where Lumen helps with citations in AI presentations
Lumen does not turn AI into a primary source. It helps teams keep the deck workflow closer to the evidence workflow: create from a topic, notes, briefs, articles, and uploaded reference material, review the outline, edit the slides, and keep citations, provenance, and human source review close to the final presentation.
Decks can cite uploaded source material and approved web research, with a built-in citation, trust-check, and source-review surface for reviewers before handoff. That matters because citations are easiest to lose when a deck is copied between tools, pasted into templates, or manually rebuilt under deadline pressure.
The current numeric check compares generated slide and Stanza scene prose with uploaded material and captured web search results. It can correct or remove an unsupported figure; it does not verify all claims, dates, quotations, or chart values. A search-result snippet and a saved source record are review aids, so open the original source and confirm its date, scope, unit, and meaning before citing it.
For a responsive Stanza story, keep the citation close to the scene that uses it and inspect the reading view your audience will receive. A deck-to-story conversion reorganizes the material; the reviewer still owns the claim and its supporting source.
FAQ
Can I cite ChatGPT, Claude, Gemini, or another AI model as the source?
You can disclose that the model assisted the deck, but factual claims should cite the underlying source. The AI model is a tool in the workflow, not the evidence for a business fact.
Do I need to disclose AI if I heavily edited the slides?
If AI materially shaped the deck, disclosure is usually safer. Editing changes the final authorship mix, but it does not erase the fact that AI assisted the workflow.
Should citations appear on every slide?
Not always. Put citations where claims need support. A methodology or appendix slide can work for some decks, but high-stakes factual claims should be easy to trace from the slide itself.
How do you cite a PPT you made with ChatGPT?
Disclose the assistance, then cite the underlying sources separately. A workable pattern: a short methodology or speaker note naming what the model did (outline, drafting, summarisation, image generation) and who reviewed it, plus a real citation attached to every factual claim on the slide that carries it. Do not list the chat transcript as the source of a market size or a quote — find and verify the primary source, because a plausible sentence is not evidence.
Do you have to disclose AI in a business presentation?
There is no single rule, so match the audience's standard rather than guessing. Internal brainstorms usually need nothing. Board decks, investor decks, regulated and compliance material, client deliverables, and academic submissions typically expect an explicit statement of what AI contributed and who is accountable for the result. Check your own organisation's policy and the recipient's before deciding, and when the two disagree, follow the stricter one.
Do AI-generated images in a presentation need attribution?
Treat them as illustration, not evidence, and label them accordingly. Note that the visual was AI-generated where your policy or audience requires it, and never let a generated image stand in for a real chart, a real screenshot, or a real photograph of a product or person. The failure mode to avoid is a generated visual being read as documentation of something that happened.
Updated
More from Gixo