How to Use AI for Slide Decks: It Builds the Deck, Not the Argument
Ask any AI tool for “a deck on our Q3 results” and you will have twelve slides in about twenty seconds. Clean layout, consistent fonts, sensible section breaks, a title slide you did not have to think about.
Then you read them and there is nothing there. Every headline is a noun phrase. Every slide is four bullets. It looks like a presentation in the way a stock photo of a meeting looks like a meeting.
The failure is specific and so is the fix, and the fix takes about ten minutes.
Start with what it genuinely does well
This is worth being fair about, because the time saved is real.
AI is good at the parts of deck-building that are mechanical: turning a document you have already written into a slide sequence, keeping formatting consistent, suggesting where a chart would go, writing speaker notes, cutting twenty slides to twelve, and producing three versions of an opening so you can pick one. All of that is genuine work that used to eat an afternoon.
What it cannot do is decide what you are trying to say. And on a slide, that turns out to be almost the whole job.
Why default AI slide decks fail
Ask for slides and you get topic headlines with bullets underneath, because that is what almost every deck in the training data looks like. It is also the shape the research says performs worst.
Joanna Garner and Michael Alley at Penn State showed a technical presentation to 110 students in two versions. Same content, same speaker. One version used PowerPoint’s ordinary defaults: a topic phrase at the top, bullets below. The other used an A slide structure where the headline is a full sentence stating the point of the slide, and the body is one piece of visual evidence supporting it: a chart, a photo, a diagram. No bullet lists. Instead of a slide headed “Q3 Pipeline” with four bullets, you get one headed “Q3 pipeline grew 40%, but all of the growth came from one account,” with the chart that shows it. structure.
The assertion-evidence group came out ahead on comprehension, held fewer misconceptions about the material, reported lower mental effort, and still recalled more when tested later. Related work found people are more likely to remember a point when it sits in the sentence headline than when it sits in a bullet underneath.
There is a second reason bullets hurt, and it explains the thing where you watch a presenter read their own slides. Richard Mayer’s work on multimedia learning includes the The finding that people learn better from narration plus a visual than from narration plus a visual plus the same words written on screen. The text you think is reinforcing your point is competing with you for the audience’s attention, because they cannot read and listen at full capacity at once. . Your bullets are not backup. They are interference.
So the default AI deck is not merely bland. It is built in the shape that measurably works least well.
The same slide, two ways
A topic is not a point
Topic headline (what AI gives you)
Could be a folder name. The audience leaves knowing the subject and not the point.
Assertion and evidence
The headline carries the claim. The visual is the proof, not a repeat of the words.
Shown the same talk both ways, 110 students given the assertion-evidence version understood more, held fewer misconceptions, reported less mental effort, and remembered more when tested later.
Write the assertions. That is the actual deck.
Here is the ten minutes that changes the output. Before you open any AI tool, write one sentence per slide. Each sentence states the point of that slide, not its subject.
The test is whether the headline could be a folder name. “Market overview” could be a folder. “We are losing share in the only segment that is growing” could not.
- Budget becomes “We can fund this within the existing budget if we delay the hardware refresh.”
- Customer feedback becomes “Customers like the product and cannot work out how to buy it.”
- Next steps becomes “We need a decision on pricing by the 30th or we miss the quarter.”
Now read your sentences in order, ignoring everything else. That list is your argument. If it does not hold together as a story, no amount of design will save the deck, and this is exactly the failure that Workslop: The Polished Report That Wastes Everyone’s Time describes: output that looks finished and transfers no meaning.
This step is yours because it depends on what you have decided, what your audience already believes, and what you want to happen next. The model has none of that, which is the same reason giving it real context beats describing the topic.
Then hand the rest over
With the assertions written, AI does the remaining work well. Paste them in and ask for something like this:
Here are the assertions for a presentation, one per slide, in order. For each one: suggest the single strongest piece of visual evidence (chart type and what it plots, a diagram, or a photo), keep the assertion as the headline exactly as written, and add two sentences of speaker notes saying what I should say out loud. No bullet lists. Tell me which assertions you think are weak or unsupported.
That last clause earns its place. The model will often flag the slide where you are asserting something you have no evidence for, which is the slide that was going to get you a difficult question.
From there the follow-ups are the mechanical work: “cut this to eight slides and tell me what I lost,” “give me three openings,” “rewrite the speaker notes for a skeptical finance audience.” If you build decks to the same shape often, turn the whole thing into a reusable setup rather than retyping it, which Build a Reusable AI Assistant covers.
Check three things before you send it
- Every headline makes a claim. Any headline that could be a folder name goes back for a rewrite. This is the one check that does most of the work.
- You are not reading the slides aloud. If your speaker notes duplicate what is on screen, cut one of them. Usually the screen.
- Every number came from you. AI will happily produce a plausible market size or growth rate to fill a chart it thinks belongs there. Anything numeric in the deck needs to trace back to a source you have, and How to Check AI’s Work covers what that check has to catch.
The takeaway
The reason AI decks feel empty is not that the writing is bland. It is that the model defaulted to topic headlines and bullets, which is the structure that tests worst for comprehension and recall, and it did so because you did not tell it what each slide was supposed to say.
Write one sentence per slide, stating the point, before you open the tool. Ten minutes of that turns AI from a thing that generates decks into a thing that builds the one you already decided on.
Sources: the comprehension, misconception, cognitive-load and delayed-recall findings are from How the Design of Presentation Slides Affects Audience Comprehension: A Case for the Assertion-Evidence Approach (opens in a new tab) (Garner and Alley, International Journal of Engineering Education, 29(6), 2013). Worth noting the scope: 110 engineering students watching a technical presentation, which is not the same audience as a boardroom, though the mechanism it tests is not specific to engineering. The redundancy principle is from Richard Mayer’s body of work on multimedia learning. The folder-name test and the prompt above are mine, built on the assertion-evidence structure rather than measured separately.
Related: Workslop: The Polished Report That Wastes Everyone’s Time is the same failure in prose, and worth reading if your decks look finished and land badly. Giving AI Context: The Skill That Actually Makes It Useful is why supplying the assertions works so much better than describing the topic.