A presentation that explains AI works best when each slide carries one idea, the speaker says what the slide shows instead of reading it, and every text block meets the 4.5:1 contrast ratio set by the Web Content Accessibility Guidelines. A 2001 study by Richard Mayer and colleagues found that students who saw a narrated animation with duplicated on-screen text scored lower on retention and transfer than students who saw no text. The same logic applies to a PowerPoint deck about neural networks or chatbots: fewer words on screen, more explanation in the room.
Key takeaways
- In four experiments with college students, adding on-screen text that repeated or summarized narration lowered retention and transfer scores compared with no text.
- Interesting but irrelevant details, such as extra anecdotes or video clips, also lowered transfer in the same research.
- WCAG 2.2 level AA asks for a 4.5:1 contrast ratio for normal text and 3:1 for large text, meaning 18 point or larger, or 14 point bold.
- The Section 508 program points presenters to slide masters and templates, which reduce common accessibility errors, and to written alternative text for images.
- Any claim about a specific AI product needs a date on the slide, because versions, prices, and features change.
What the research on cognitive load says about slides
Mayer, Heiser, and Lonn published "Cognitive constraints on multimedia learning: When presenting more material results in less understanding" in the Journal of Educational Psychology in 2001. College students watched an animation of how lightning forms, with narration. In the first two experiments, some viewers also saw on-screen text that summarized or duplicated the narration. Those viewers scored lower on later tests of retention and of transfer, which means applying the idea to a new problem, than viewers who saw no text. The authors read this as a redundancy effect: the eyes have to split between the picture and the words.
The same paper reports a second result. Adding interesting but irrelevant material, for example extra sentences in the narration or unrelated video clips, reduced transfer. A later review of that line of work is cited as finding the same benefit of concise presentations in thirteen of fourteen tests. The limits are plain, though. The participants were college students, the topic was one science process, and the format was a short computer lesson. A conference talk to a mixed audience is a different setting, so the finding is a reason to test slides, not a law.
A slide sequence for explaining an AI concept
The table is an illustrative outline for a ten-minute talk on any AI system. The slide titles are suggestions and the checks are the useful part.
| Slide | Question it answers | What goes on the slide | Check before presenting |
|---|---|---|---|
| 1 | What task does the system perform? | One sentence and one example input and output | Could a listener who sees only this slide state the task? |
| 2 | What data does it learn from? | A sample of the data, not a description of it | Is the source named and dated? |
| 3 | What is training and what is use? | Two panels: learning from examples, then answering new inputs | Are both labeled in plain words? |
| 4 | What does a good result look like? | One real output | Is it from a documented source, not a mock-up? |
| 5 | Where does it fail? | One real error and what caused it | Is the error shown as often as the success? |
| 6 | Who sets the rules? | The regulator or standard that applies, with a date | Is it enacted, voluntary, or proposed? |
| 7 | Where can the audience check? | Two or three primary sources | Does each link open? |
Slide 6 deserves care. The U.S. National Institute of Standards and Technology released its AI Risk Management Framework 1.0 on January 26, 2023, and describes it as intended for voluntary use. A deck that calls it a requirement misstates it.
Dating the claims
AI slides age fast. A statement such as "the latest model can do X" is wrong the moment a newer release ships. Giving every product claim an "as of" date, and naming the model version, keeps a deck honest for the next presenter who reuses it. The article on model versions and dates linked below shows how quickly a version comparison can go stale.
Contrast, templates, and alternative text
Contrast is a practical matter in a lit room before it is a compliance one. The WCAG 2.2 Understanding page for Contrast (Minimum) sets level AA at 4.5:1 for normal text and 3:1 for large text, where large means 18 point or larger, or 14 point if bold. A free contrast checker can measure the ratio for any color pair before the deck is finalized.
The U.S. government's Section 508 program guidance on presentations says the PowerPoint slide master and templates can reduce common accessibility problems and directs authors to the W3C Alt Decision Tree for describing images. For a diagram of a neural network, a useful alternative text states what the diagram claims, for example "three layers of nodes, each node linked to every node in the next layer", and does not repeat the caption. The page is written for presentations that must conform to the Section 508 Standards, and speakers outside that setting can use it as a ready benchmark.
A pre-talk checklist
- Cut each slide to one claim and delete anything the speaker will say aloud word for word.
- Remove decorative images, stock robots, and anecdotes that do not support the claim.
- Replace any slide of dense text with a figure plus a short spoken explanation.
- Test the deck on a classroom projector or a laptop at arm's length.
- Run the built-in accessibility checker and read the slides in reading order.
- Put the date and version beside every product claim.
- Close with sources the audience can open, and keep that slide on screen during questions.
This article gives general information about presentation design and accessibility standards. It is not legal advice on accessibility obligations, which depend on the organization and jurisdiction.
FAQ
How many words should an AI slide contain?
No source sets a number. The research above supports keeping text short and not repeating what the speaker says, so a headline sentence and a figure is a workable starting point.
Should the deck explain how the math works?
Only if the audience needs it for a decision. For most general audiences, a worked example with real input and output explains more than equations, and the equations can sit in an appendix.
Are AI-generated slides acceptable?
Drafts can save time, but every factual claim and number needs checking against a source before the deck is shown, because generated text can include errors that read confidently.
Is a slide deck a substitute for a handout?
The two do different jobs. Slides support a live explanation, while a handout or notes page can hold the detail and the full source list for reading later.
Related reading
- Explaining versus predicting in data work
- Why model versions and dates matter
- A dated timeline of Baidu's Ernie Bot
- Measuring how close a route is to optimal
More in AI Basics.
Last verified: October 2026
Sources
- Mayer RE, Heiser J, Lonn S., Cognitive constraints on multimedia learning: When presenting more material results in less understanding, Journal of Educational Psychology 93(1), 2001
- Mayer RE, Moreno R., Nine ways to reduce cognitive load in multimedia learning, Educational Psychologist 38(1), 2003
- Understanding Success Criterion 1.4.3: Contrast (Minimum), W3C Web Accessibility Initiative, WCAG 2.2
- Create accessible presentations, U.S. General Services Administration
- AI Risk Management Framework, National Institute of Standards and Technology, 2023