ChatGPT designs decks now. Here’s what it gets right, and where it still needs rescuing

Something shifted this month. When OpenAI announced GPT-6 in early September, presentations were in the headline claims. Their words, not mine: it’s their best model yet for adhering to existing templates and producing slides that are well laid out and convey key points with a structured narrative. That follows the official ChatGPT add-in for PowerPoint from May, which builds and edits decks inside the app itself. The biggest AI product on earth has decided slides are worth being good at.

I design presentations for a living, and I’ve spent the last year fixing the ones AI made. So I’m not going to tell you this doesn’t matter. It matters a lot. I’m going to tell you what actually changes, what doesn’t, and how to know which deck you’re holding.

What genuinely got better

Credit where it’s due. Three things are now real.

Structure. Give it a topic, an audience and a slide count and it will produce a logical arc with a beginning, a middle and a call to action. For most people that’s the hardest part of starting a deck, and it’s now the easiest.

Writing. This is where the tools earn their keep. Turning a forty-page report into slide-sized points, drafting speaker notes that don’t just repeat the bullets, tightening copy, restructuring an argument that isn’t landing. I use it for this daily and I’d be slower without it.

Template adherence, on paper. This is the new claim and the interesting one. Previous generations of these tools ignored your template and improvised. If GPT-6 genuinely respects an existing master, that’s a real step, and it moves the problem somewhere more interesting. I’ll come back to that.

And it isn’t only OpenAI. Claude has been building decks inside PowerPoint since last autumn, and Gemini does the same inside Google Slides. Whichever tool your organisation has standardised on, the slide-making is now in the box. That’s worth saying plainly, because it means what follows isn’t a ChatGPT problem. It’s a generated-deck problem, and every one of these tools has it.

I’ve compared all the main tools side by side here.

What a better model doesn’t fix

Here’s the thing the announcement doesn’t say and the tool vendors quietly admit in their own guides: a better language model does not make better-looking slides. Layout, typography, hierarchy and visual judgement aren’t language problems. The model decides what to say and in what order. Something else decides how it looks, and that something is either a template, a designer, or luck.

So the failures I see on my desk haven’t gone away. They’ve just got faster to produce.

Formatted, not designed. The content lands evenly and competently. Every slide gets the same weight and rhythm. Design is the opposite: deciding what matters on this slide and making it unmissable. The tools don’t know which number is the one the room came for.

The template gets obeyed literally, not intelligently. Adherence means the fonts and colours follow the master. It doesn’t mean the tool knows a nine-bullet slide should be three slides, or that the chart needs to be the slide, not a decoration on it. A template followed without judgement produces something brand-compliant and dull.

Things go missing. Squeeze content into layouts and footnotes vanish, sources drop off charts, the last two bullets get cropped. Nobody notices until someone senior does.

Everything looks the same. AI decks have a look now, and audiences are learning to spot it. When your proposal resembles the other side’s proposal because you both pressed the same button, whatever made you distinctive has left the room.

The judgement layer is empty. What deserves a full slide. What gets cut. Where the argument needs to slow down. What this specific room, on this specific day, needs to see first. No prompt contains that, because you’d have to already know the answer to write it.

A brief landed in my inbox recently with a note against one slide that said, in full, “AI did this, fix it.” That sentence is the whole market in six words. The person who wrote it knew exactly what was wrong. They just couldn’t make it right.

The problem is moving upstream, too. Another recent brief arrived as a large rectangle filled with a paragraph, clearly generated, explaining how the slide should be designed: three boxes here, this content in that one, and so on for two hundred words. It took me far longer to decode than a rough sketch the briefer could have scribbled in a minute, the kind I’d have understood in seconds. Prose is a terrible way to describe a picture, and the tools produce prose. When the deciphering is billable, the AI brief that was meant to save time ends up costing more than the ugly drawing would have.

The interesting consequence: your template just became the last line of defence

If the model now follows your template, then your template is doing the brand’s work every time anyone generates anything. Which means the quality of the template matters more than it did last year, not less.

Most corporate templates were built before anyone generated into them. They have hardcoded colours the tools can’t inherit, fonts set per slide instead of at theme level, layouts named things no machine could map, and no library of finished slides for anything to draw from. Point a template-adhering AI at a template like that and it will adhere to the mess.

An AI-ready template is mostly just a properly built template, held to a higher standard because the audience is now partly software. True theme colours. Theme-level fonts. Clean, named, semantic layouts. A slide library people and tools can pull from instead of improvising. I’ve spent years building templates that survive thousands of users across a network. Surviving AI is the same discipline, and it’s now the thing I’d fix first in most organisations.

I wrote more about this on the AI presentation design page, including what I do when a template has already been overrun.

How to know which deck you’re holding

A simple test. Ask what the deck is for.

If it’s the Tuesday catch-up, the internal update, the first draft for a conversation that will change everything anyway, the generated deck is fine. Genuinely fine. Use the tool, save the hours, move on.

If it’s the pitch, the board, the investor meeting, the client who has seen a hundred decks and will judge yours in the first ten seconds, then generated isn’t enough. Not because the tool is bad, but because the tool can’t know what that room needs. That’s a design decision, and it’s the one that decides whether the deck circulates or wins.

The expensive mistake isn’t using AI. It’s not knowing which deck is which.

Where I’ve landed

Use these tools. Use them for structure, for drafting, for speaker notes, for turning research into something slide-shaped. They’ve moved the starting line a long way, and I’d be a fool to pretend otherwise.

Then look honestly at what the meeting is worth. If it’s worth more than the afternoon a generated deck saves you, bring in the judgement the software can’t have. Sometimes that’s a fix-up of what the tool produced. Sometimes it’s a template built so the tool produces something better next time. Occasionally it’s a designer from the start.

If an AI tool made your deck and it shows, that’s now its own service. If it’s the template that’s been letting the tools down, that’s a different fix, and usually a more permanent one.

Either way, the deck that matters is still the one somebody designed.

Dan Plumb is a presentation designer with sixteen years of agency-grade work, formerly Global Head of Design at Publicis Media, working with clients directly at depicts.com.