What an audience notices within five minutes. These are the failures we keep finding while checking 167 layouts, and each one has a different fix.
Most AI-generated decks are not badly made. The colors are fine and everything lines up. Then, around the third slide, people start to feel that they have seen this before and that it does not belong to this company. That feeling comes from patterns, not taste. The seven below are the ones we keep running into, and they have different causes and different fixes. None of them require a designer to fix. Most are decisions made in the wrong order, which is why they repeat across otherwise unrelated decks.
| Tell | Cause | Fix |
|---|---|---|
| Everything centered | No hierarchy was decided | Make one element the biggest |
| Always three cards | Form chosen before content | Count the items, then pick a form |
| Titles as sentences | The script pasted onto the slide | Noun phrase, number first |
| Numbers with no source | Plausible standing in for true | Year and source on the same slide |
| Decoration with no job | Filling empty space | One accent, used once |
| Clipped or overflowing text | Never rendered at full size | Open it at presentation size |
| Spacing drifts per slide | Values decided slide by slide | Move spacing and sizes into variables |
Centered text never tells the reader where to start.
The cover is centered, the body slides are centered, the closing is centered. This is the first tell. Centered text gives the reader no instruction about where to start. When every element carries the same weight there is no hierarchy, and without hierarchy the audience has to decide for itself what matters. A cover with one line belongs in the center. Once there are two or more blocks of information, left alignment plus a size difference reads better. The fix is small. Pick the one element that should be read first and make everything else smaller than it. Once that ranking exists, the audience stops scanning and starts reading, and the presenter gets to talk over a slide instead of against it.
Counting the items first solves half of it.
Two items become three cards. Five items become three cards. The extra space gets padded and the extra items get dropped. The content should decide the form, and here the order is reversed. Slides with three cards are overwhelmingly common in training data, so that is what comes back. Count the items first, then choose an arrangement built for that count. Two items read better as a contrast, five read better as a list than as a grid, and the comparison and feature layouts are split by count. When the number of items changes, the arrangement has to change too. Forcing them into the same grid is what breaks the spacing. A quick test: cover the slide and say how many things it is comparing. If the answer is not the number of boxes on screen, the form is wrong.
A title is a half-second label, not the script.
"We improved customer satisfaction through a data-driven approach" is not a slide title, it is the script. A title is a label the audience reads in half a second. As soon as it becomes a sentence it takes time to read, and it competes with the person speaking. Cut it to a noun phrase and lead with the number: "Satisfaction up 32%". The reasoning is what the presenter is there for. If you want a rule of thumb, keep it near 40 characters. Longer than that usually means two messages are sharing one slide. Splitting those two messages into two slides costs nothing. Slide count is not what makes a deck feel long; unread slides are.
Not this
We improved customer satisfaction through a data-driven approach
This
Satisfaction up 32%
A wrong number costs more than an empty space.
A market size sits in 90-point type with no year attached and no idea who produced it. Models are good at filling in plausible figures and equally good at leaving out where they came from. In front of an investor, that single slide costs the deck its credibility. If a number is on the slide, put the year and the source on the same slide, even in small type. Cutting it for space also cuts the answer to the first question you will be asked. A figure you cannot verify is better left out. A blank space is cheaper than a wrong number. This is also the failure that survives editing, because a number that looks confident rarely gets questioned by the person who wrote it.
Not this
A $90B market
This
US mid-market segment, $4.2B (IDC, 2025)
When everything is emphasized, nothing is.
A purple-to-blue gradient behind the text, an icon on every bullet. None of it carries information and all of it takes room. Decoration exists to pull the eye somewhere. When everything is decorated, nothing pulls. Use icons only when they distinguish categories. Five items with five different icons is noise, not distinction. Color works the same way. Choose one accent and spend it once per slide, and that one use actually reads as emphasis. Gradients deserve a separate warning. They shift the background value behind text, so contrast stops being predictable and small type suffers first.
This is a verification problem, not a design one.
A title wraps to two lines and pushes the cards below it, or a description runs past the edge of its box. This is a verification problem rather than a design one. A generated slide has usually never been rendered at full size. It is fine as code and broken on screen. Before a layout goes into our catalog we check five things automatically: content outside the canvas, text clipped by its container, content running into the footer, elements overlapping that should not, and elements hidden behind others. If you build slides yourself, at least open them once at the resolution you will present at. That is where most of the "it looked fine on my laptop" stories come from. Projectors and shared screens crop differently too, so the edge of a slide is the least safe place to put anything that has to be read.
Nobody names this one. They remember the deck as careless.
This one is the hardest to see and the one that lingers. Slide three and slide seven have different margins, and body text shifts size along the way. It follows from building one slide at a time. Every decision made fresh produces a slightly different value. The fix is to stop deciding per slide. Pull spacing and type sizes into variables and the whole deck keeps one rhythm. Audiences never name this problem. They remember it as the deck feeling a little careless. It also compounds. Ten slides with small inconsistencies read worse than three with the same flaw, because the eye starts predicting and keeps being wrong.
Fixing all seven one at a time is slower than changing how the deck gets made. Write the one thing this slide has to say, count the items, choose an arrangement that fits that count, then write the text to the space it has. The usual order is the opposite. People write everything, look for a layout, and shrink the text or stretch the box when it does not fit. Numbers one through seven are produced by exactly that sequence. That is the reason getdeck exists. When a verified arrangement comes first, five of the items on this list never get the chance to happen.