Lesson 5 of 7 · 7 min read

The AI tells

Check an AI draft for the structure, surface and claim signs of a deck nobody checked, and fix the claim tells first.

What an AI tell is

An AI tell is a sign that a deck came out of an assistant and nobody checked it. The tells fall into three groups: structure, surface and claims. Checking for them is the fourth rule of this course.

Lessons 1 to 4 built your argument by hand. From here on, an assistant does more of the building. This lesson shows what an assistant tends to produce when it gets a loose request, so you can spot it in a draft. Lesson 6 then builds the deck in a way that avoids most of it.

The checklist rests on expert consensus. Practitioners who build and review decks agree on the same short list. No source counted how often each tell appears. No study showed that audiences detect AI-made slides, or engage less with them.

Most of the sources sell presentation tools. Perceptis, Plus AI and Pekto each publish a list of tells and a product that claims to fix them. A developer writing on DEV Community names the same items, and so do Anthropic’s own design instructions for Claude. Use their checklists and ignore their tool rankings.

Structure tells

Structure tells come from the tool’s defaults, not from your argument. The most common is one layout repeated on every slide. Another is a title over three equal cards, each with a line icon. A third is a deck where every slide gets the same visual weight. A fourth is a layout centered and mirrored on every slide.

These tells matter because layout should follow the claim. Lesson 3 gave every slide one claim. Some claims carry the decision, and their slides deserve more space and a larger chart. A deck where every slide looks the same hides which claim matters most.

Pekto’s advice is to match the structure to the data. A roadmap suits a timeline. A comparison of competitors suits a grid. Three equal cards suit three equal things, which a real argument rarely has.

Surface tells

Surface tells are the look of the slides. The sources name default fonts, with Inter the example most often given. They name default color schemes: a purple gradient on white, or a cream background with muted red accents.

Plus AI adds two more. One is a colored accent bar with a small all-caps label above the title. The other is an italic serif phrase at the end of a sans-serif title. Perceptis adds icons that mean nothing and glossy stock 3D images.

Anthropic’s frontend-design skill tells Claude itself to avoid Inter, Roboto, Arial and system fonts, purple gradients on white, and predictable layouts. The model’s maker treats these as defaults to override.

Perceptis notes that better prompts fix content tells but not a tool’s built-in visual defaults. Your own template fixes most surface tells. Lesson 6 builds the deck inside it and turns the rest into written brand rules.

Claim tells matter most

Claim tells are problems with what the slides say. Topic-label titles such as “Q3 Results” state no claim. Bullets restate the obvious. Numbers appear with no source. Plus AI warns that stat tiles get filled with plausible round numbers.

A deck with a default font can still persuade. A deck with an invented number fails, and it fails in front of the people you need. Fix claim tells first, every time.

Claim tells also move work onto your audience. One Hacker News user’s boss sent an AI-made deck for review. It was mostly right but had misleading and wrong points. The reviewer expected hours of fact-checking, against a few minutes for a deck a person had made.

Another user described polished AI decks that management liked. When the user asked what a term on a slide meant, the presenters could not answer, because the model had chosen it. These are single reports from a technical forum. They show how the failure looks, not how often it happens.

Run the check in a fixed order

Check the three groups in order of harm: claims first, then structure, then surface. A claim tell can mislead the reader. A structure tell hides which claim matters. A surface tell changes how the deck looks, not what it says.

Read the claims with your sources open beside the deck. Trace every number, date and named fact to a line in a source. Mark each estimate as an estimate, and say whose it is. Delete any number you cannot trace, even if it looks right.

Then read the titles alone, in order, as Lesson 3 taught. Next, flip through the slides without reading them, and note which layouts repeat. Look at the surface last. Most surface tells disappear once the deck sits in your own template.

Tells describe patterns, not authorship

Wikipedia’s editor guide to signs of AI writing gives a caution that applies to slides. Its signs describe patterns, and they do not prove who wrote a text. The guide cites studies in which people barely beat chance at spotting AI text.

So this check is not about hiding that you used AI. A deck made entirely by hand can carry every tell on the list. The aim is a deck that makes its point with numbers you can trace.

The same guide lists writing signs worth checking in slide text and speaker notes. Examples are words such as “delve” and “pivotal”, “serves as” where “is” would do, and bulleted lists with a bold label on every line.

Worked example · Larkfield Supply

Morgan reads the first AI draft

Before building the real deck, Morgan tests what an assistant does with a loose request. Morgan pastes in the one sentence from Lesson 2 and last peak’s notes, and asks for a seven-slide deck. There is no template and no slide brief.

  1. Morgan checks structure first. Five of the seven slides use the same layout: a title over three equal cards, each with a line icon. The cost slide gets the same space as the agenda slide.
  2. Morgan checks the surface. The title slide has a purple gradient. Every slide uses the tool’s default font, with a small all-caps label above the title. A stock 3D image of boxes on a conveyor fills half of one slide.
  3. Morgan checks the titles. They read “Background”, “The Challenge”, “Our Solution”, “Financial Impact” and “Next Steps”. None of them states a claim, so the titles-only test from Lesson 3 fails at once.
    What would you do next?Morgan has found tells in all three groups. Which group should Morgan fix first?Think it through, then show the answer

    Claims. The template in Lesson 6 will fix most surface tells, and the slide brief will fix the layouts. A wrong number or a missing claim can mislead Priya, who reads every figure the day before.

  4. Morgan traces every number to the notes. The 97% and 88% on-time figures match. A stat tile reading “30% faster order processing” appears in no source. Another slide says on-time shipments “will return to 98%”.That answers the question you guessed at: An AI draft shows a stat tile reading “30% faster order processing.” Guess: what should you do before you keep it?
    What would you do next?Morgan’s own estimate is about 96%. What should the slide say?Think it through, then show the answer

    It should give about 96%, and it should say the figure is Morgan’s own estimate, not a measured result. The label belongs in the title or directly under the chart, where Priya cannot miss it.

  5. The customer slide says “customer satisfaction declined significantly”. The source is narrower: two of the three largest customers complained in writing. Morgan replaces the vague claim with that fact.

Morgan’s list holds three structure tells, four surface tells and eight claim tells: five topic-label titles, two numbers with no source and one vague claim. Morgan does not fix this draft slide by hand. The claim tells came from asking for slides before writing a brief, so Lesson 6 rebuilds the deck from an approved brief in Larkfield’s template. Morgan keeps the list for the critic pass in Lesson 7.

Starter prompts

  • Find the AI tells in my deck

    Here is my deck, slide by slide, with titles, body text and chart labels:

[paste your slide text, or attach the deck as a PDF]

Check it for AI tells in three groups.
Structure: one layout repeated on most slides; a title over three equal cards with icons; every slide given the same weight; centered, mirrored layouts on every slide.
Surface: default fonts; default color schemes such as a purple gradient on white; accent bars with small all-caps labels; an italic phrase ending a title; icons or stock 3D images that carry no information.
Claims: titles that name a topic instead of stating a claim; bullets that restate the obvious; numbers with no source; round numbers in stat tiles.
For each tell, give the slide number, the group and what you saw. Do not rewrite the slides. Do not guess whether AI made the deck. If you cannot see the visual design from what I gave you, say which checks you could not run.

  • Trace every number to a source

    Here is the text of my slides:

[paste your slide text]

Here are my sources:

[paste or attach your sources]

List every number, date and named fact on the slides. For each one, quote the line in my sources that supports it, or write "no source found". Mark any figure that is an estimate or a forecast, and say whether the slide labels it as one. Do not invent sources. Do not correct any number yourself.

When not to use AI

  • Deciding whether a colleague used AI. The tells describe patterns and prove nothing about who made a deck.
  • Confirming that a number is right. An assistant can list every number on your slides, but only the source can confirm it. Check each one yourself.
Recall from memoryName the three groups of AI tells, and say which group you fix first and why.Answer it without looking back, then show the answer

Structure, surface and claims. Fix claims first, because a claim tell can mislead the reader. A structure tell hides which claim matters, and a surface tell changes only how the deck looks.

Exercise · your project

Check your own deck for AI tells, claims first.

  1. If you have no draft yet, ask your assistant for a quick one from your one sentence and your notes, so you can see its defaults.
  2. Run the first prompt and list each tell under structure, surface or claims.
  3. Run the second prompt with your sources, and fix or remove every number that has no source.
  4. Label every estimate on the slides as yours.
  5. Save the list of tells. Lesson 6 turns the surface tells into brand rules, and Lesson 7 uses the whole list in the critic pass.

Sources