How often each tell showed up
| Tell | Decks with it | Slides with it |
|---|---|---|
| The accent color means two things | 16% | |
| Too many levels of text | 15% | |
| The slide says it twice | 11% | |
| A subtitle that only repeats the headline | 6% | |
| Everything is the same size | 4% | |
| Labels that name the obvious | 4% | |
| Headlines built from fragments | 4% | |
| Buttons nobody can press | 3% | |
| The words don't match the screenshot | 1% | |
| A second headline | 1% |
Check your own deck
Open your deck, go slide by slide, and tick each tell you find at least once.
0 of 10. Tick the tells your deck has.
1 The accent color means two things
In 63% of decks and 16% of slides.
An agent reaches for the accent wherever something feels important, so the headline promise and the problem in the diagram end up the same color. The eye can't tell which one is the point.
Fix it. Give the accent one job per slide. Everything else goes neutral.
Instead ofCoral on "saves 4 hours" in the headline, and coral on the red-flag step in the diagram.
TryCoral on "saves 4 hours" only. The diagram step is dark gray with a label.
2 Too many levels of text
In 62% of decks and 15% of slides.
Agents explain. A slide gets a headline, a subhead, card titles, card body and a caption, and the reader has five things to read before they know what to look at.
Fix it. Keep the headline, one focal element and one quiet line. Move the rest to speaker notes.
Instead ofA chart, its finding as a subhead, a sentence explaining the chart, and three labeled callouts.
TryThe finding as the headline, and the chart.
3 The slide says it twice
In 55% of decks and 11% of slides.
Generated copy pads. The card repeats the subhead, and the summary line repeats the card, so the slide reads like it doesn't trust you to get it the first time.
Fix it. Say each idea once, in the place it does the most work.
Instead ofA card titled "Faster onboarding" with the body "New hires get set up faster."
TryOne line: "New hires are set up on day one."
4 A subtitle that only repeats the headline
In 34% of decks and 6% of slides.
Headline plus subtitle is the default shape of a generated slide, even when the subtitle has nothing to add.
Fix it. Cut it, or make it carry a fact the headline doesn't: a number, a date, who it's for, a source.
Instead of"Churn fell to 3% in Q3" over "3% churn".
Try"Churn fell to 3% in Q3" over "Down from 7% after the onboarding change."
5 Everything is the same size
In 28% of decks and 4% of slides.
Grids of equal cards are the easiest layout to generate. They are also the layout that makes no argument.
Fix it. Decide which item is the point and make it bigger, darker or alone. Shrink the rest.
Instead ofFour stat cards at the same size: 40%, 12 days, 3x, 98%.
Try40% large, with the other three in one quiet row under it.
6 Labels that name the obvious
In 23% of decks and 4% of slides.
Eyebrows like THE WORKFLOW over a list of steps look designed but tell the reader nothing.
Fix it. Delete the label, or replace it with something the reader doesn't know yet, like a step count or a date.
Instead ofTHE PROCESS, above a list of steps.
TryNo label. Or "FOUR STEPS, ONE WEEK".
7 Headlines built from fragments
In 22% of decks and 4% of slides.
Clipped fragments read like ad copy, and readers notice the rhythm.
Fix it. Write the sentence you would say out loud.
Instead of"Built for finance. Loved by auditors."
Try"Finance teams close the month two days sooner, and auditors get every receipt."
8 Buttons nobody can press
In 18% of decks and 3% of slides.
Agents mock up interfaces. A row of pills with one selected implies the other options matter, but nobody can click a slide.
Fix it. Show the state that matters as a plain fact, or show every state side by side.
Instead ofPills for "Monthly", "Quarterly" and "Annual", with "Monthly" highlighted.
TryOne line: "Billed monthly."
9 The words don't match the screenshot
In 9% of decks and 1% of slides.
The copy is written from the brief and the screenshot is taken from the product, so they drift apart. Anyone who knows the product sees it at once.
Fix it. Use the product's own words, exactly as they appear on screen.
Instead ofA callout reads "Smart routing" next to a screen whose menu says "Rules".
TryThe callout reads "Rules".
10 A second headline
In 9% of decks and 1% of slides.
Two lines at headline size compete, and the reader has to choose where to start.
Fix it. One headline per slide. Everything else gets visibly smaller.
Instead of"Pricing by region" as large as the headline above it.
Try"Pricing by region" as a small label over the table.
The tells everyone knows were the rarest
Our agents are told to avoid these, and code checks every slide for them, so few slipped through. Banning a word works. Design judgment is harder to instruct.
- Middle-dot separators"Fast · Simple · Secure"6% of decks
- Em dashes"The results were clear — and fast"3% of decks
- Words like unlock, seamless, elevate"Unlock seamless growth"2% of decks
- "Title: subtitle" headlines"Renewals: a new approach"2% of decks
- "It's not X, it's Y""It's not a tool, it's a teammate"1% of decks
If your AI tool hasn't been told to avoid them, expect to see them. But searching a deck for dashes and dots takes a minute. The design tells above are the ones that need a second look at every slide.
How we measured
Pitch Unfairly reviews every slide of every deck before it can be published. Some rules are checked in code against the slide's text: em dashes, middle dots, "Title: subtitle" headlines, generated words, stacked fragments and type set at headline size. The rest are judged by a model that reads the slide's text or looks at its screenshot, one rule at a time.
The counts cover 1,364 deck builds and 11,319 slides reviewed from October 8 to October 22, 2026. A deck or slide counts once for a rule if the reviewer flagged it in any review round, so each figure means "showed up at least once while the deck was being built," not "was in the published deck."
Every build was made by AI agents for our own team: 803 from test briefs we use to evaluate the product, the rest real work. The agents were told to avoid all of these tells, so the numbers show what survives the instructions. The model-judged rules haven't been measured against designers' own labels yet, so read those shares as how often the reviewer raised a tell, not as a verified rate. We'll update this page as more decks are reviewed.
AI deck questions
How can you tell a presentation was made by AI?
Look at the design before the words. In 1,364 decks built by AI agents, the tells that showed up most were an accent color marking two ideas (63% of decks), too many levels of text, and the same idea said twice on one slide. Em dashes and words like "unlock" were rare.
Do em dashes mean a deck was written by AI?
Not on their own, and they are easy to remove. Our agents were told not to use them and a code check catches them, so they turned up in only 3% of decks. The tells that survive instructions are design choices, which are harder to check for.
How do I make AI-generated slides look less AI-made?
Give each slide one point: one headline, one focal element, one accent color with one job, and one quiet line of support. Then read every slide asking whether anything is said twice or labeled for no reason. The checklist on this page covers all ten tells.
Where does this data come from?
From the design reviewer in Pitch Unfairly, which checks every slide of every deck before it can be published. These are 1,364 deck builds and 11,319 slides from October 8 to October 22, 2026, all built by AI agents for our own team, including 803 built from test briefs.
Can Pitch Unfairly check my deck for these?
Every deck built with Pitch Unfairly goes through this review before it can be published, and the agent fixes what it finds. Set it up in ChatGPT, Codex or Claude Code from the setup page.
Want a deck with none of them? Build it with Pitch Unfairly or start from one of the famous decks.