I’ve read a lot of articles about everything AI fixes for designers, and I suspect a few of them were written by AI. The things it breaks get much less attention, mostly because they’re harder to spot.
The one I run into most is plausible nonsense. AI makes polished work faster than anyone can review it, and polish is very convincing. A persona with a name and a stock photo looks like research, even when the only user involved was the model. A flow can make perfect sense on its own terms and still be built on a misreading of what people are trying to do.
The annoying part is who falls for it. Juniors tend to be suspicious of AI output. Seniors on a deadline are the ones who let it skip a step they’d normally do themselves, because it looks like something they would have made.
Everything also starts to look the same. Ask for "a professional dashboard for analytics users" and you get the average of every dashboard ever made, because averages are what these models are good at. Products that feel distinct usually come from someone who knows their users well enough to break a convention on purpose. A model trained on conventions won’t do that for you. I’m happy to use it for a starting direction, and then I go somewhere else with it.
What bothers me most is research. AI can summarise thirty interviews in an afternoon, and the summary is useful. But it smooths things out. The phrase one user kept repeating that didn’t fit any theme disappears. So does the person who contradicted themselves in a way that told you more than their actual answer.
The summary keeps what doesn’t fit
Select a theme to see the quotes behind it.
"Half of March went on emails asking suppliers for the same spreadsheet."
"We send the questionnaire in January and the last answers come back in May."
"If a number is an estimate, I need to see that before it goes to the auditor."
"I can’t defend a figure I can’t trace back to an invoice."
"Exporting to Excel is the first thing I do every Monday."
"Finance wants it in their template, so I copy it across by hand."
"We keep two numbers. Ours, and the auditor’s version."
Said "the auditor’s version" six times. No one else used it.
Follow up: ask whether other teams keep a second set of figures.
"We trust the platform’s numbers completely."12 min
"I check every supplier figure against their invoice before I submit."38 min
P26 · Sustainability coordinator, logistics
Follow up: ask what they check for, and why.
Illustrative example
The summary smooths it away
Select a theme to see the quotes behind it.
"Half of March went on emails asking suppliers for the same spreadsheet."
"We send the questionnaire in January and the last answers come back in May."
"If a number is an estimate, I need to see that before it goes to the auditor."
"I can’t defend a figure I can’t trace back to an invoice."
"Exporting to Excel is the first thing I do every Monday."
"Finance wants it in their template, so I copy it across by hand."
Illustrative example
If a team only ever reads the summary, they slowly know their users less well, and nobody notices until a big decision rests on a theme that turns out to be a statistical accident. It’s a kind of debt, and it tends to come due right before a launch.
I still use AI every day. I just check its work as carefully as anything else that lands on my desk, and a bit more carefully when it sounds very sure of itself.