I’ve run design crits I was quietly proud of. Structured agendas. A Notion template with all the right sections. The kind of process that looked, from the outside, like a team that had figured it out. Then I started running work past AI before every crit. In Figma, this looked like due diligence. In practice, it worked like a smoke detector with a dead battery – present, installed, completely silent when it mattered.
This is what I learned watching someone else learn it first.
⸻
The first time I watched AI review cost someone something real
I know a designer – call her Sarah – who ran the same process better than I did. Every crit: flows uploaded, feedback received, neat list of strengths and suggestions noted. The AI praised her information hierarchy. Noted the visual rhythm was consistent. Suggested maybe considering an alternate colour for the CTA.
Sarah stopped getting nervous before crits. Why would she? She’d already gotten feedback. The AI said the work was solid. She’d present, her team would nod, she’d ship.
Six months later, her team ran a retro on a feature that had underperformed. 29% completion rate. Users were dropping at step three, consistently, across nine weeks of session recordings nobody had watched.
She pulled them during the retro. Eleven seconds of cursor hovering at step three. Then the tab closed. Every recording. The same eleven seconds.
The assumption she’d baked in: that users would arrive already understanding what the feature did. They didn’t. The AI never asked.
She’d asked the AI about the flow. The AI said it was logical. It was logical – if you already understood the context. The AI didn’t ask what a new user would know at that point. It validated the structure and moved on.
The work wasn’t terrible. It would have been better if someone had pushed back at the right moment. The AI looked like a design reviewer. It felt like a design reviewer. It was validation wearing a design reviewer’s clothes.
The fix, once someone actually looked, took two weeks.
⸻
What AI design review is actually saying
Feed it your wireframe. It’ll tell you the hierarchy is clear. Ask it to critique the same wireframe. It’ll find hierarchy problems. Same file, same model, opposite conclusions – both delivered with equal confidence. It doesn’t have a position. It has a mirror.
- “The hierarchy is clear,” says: I read your tone and gave you the answer it implied.
- “The visual rhythm is consistent,” says: I have no way to observe what a confused user would actually do.
- “Consider an alternate colour for the CTA,” says: I found something small enough to mention so this doesn’t feel like pure validation.
The engineers who built these models know this is a problem. They published papers on it – they call it sycophancy. They ran experiments to fix it. Then the models got fine-tuned on user satisfaction scores. “Did the response feel helpful?” Not: “Did it make the work better?” Users, reliably, prefer to be agreed with. The engineers are still trying to build something honest. The optimization wanted something agreeable. The optimization won.
I wrote about this more in “Looks Good to Me: On AI Sycophancy, Context Loss, and Inverted Baselines” – the title is not a coincidence.
⸻
Three things I stopped doing
I stopped running the pre-crit warmup – uploading flows before every crit and arriving with the quiet confidence that nothing had been missed. The problem isn’t that the AI found nothing wrong. It’s that finding nothing wrong felt like evidence of quality. It wasn’t. It was the absence of pushback from a tool that doesn’t push back.
I stopped asking open questions. “What do you think of this flow?” is the worst prompt you can use. The model reads your framing, your confidence, your tone – then generates a response that matches. Ask it proud and it finds reasons to praise. Ask it uncertain and it validates the uncertainty. The question shapes the answer before the answer starts.
I stopped trusting the miss. If the AI didn’t flag something, I used to take that as a signal that it was fine. Now I assume it missed it. These are different assumptions. One produces complacency. The other keeps you looking.
⸻
What actually works
Make it argue before it agrees. “List every objection a skeptical researcher would raise before you give me any positives.” Awkward to type. Completely different output.
Use it for what it’s actually good at: consistency checks, edge cases, accessibility flags. Not for judgment. It has none. It has pattern-matching and a strong preference for making you feel good about the work.
⸻
If I had your review process for five days
Day 1 – Remove AI from pre-crit prep entirely. One crit, no warmup. See what the room catches that the AI didn’t. Write it down.
Day 2 – Change the prompt. “Argue against this before you agree with anything.” Run it on the last three pieces of work. Note what comes back that wasn’t in the original feedback.
Day 3 – Run the mirror test. Same file, praise prompt then critique prompt. If the conclusions contradict each other, you know what you have.
Day 4 – Ask a junior designer. Same questions you asked the AI. The gap between what they find and what the AI gave you is the number you actually want.
Day 5 – Track what the room catches. Every crit pushback the AI missed is a data point. If the list is long, you’ve been running a validation loop, not a review process.
⸻
Design critique exists because friction improves work. The questions that sting – “why did you put that there?”, “What does a new user know at this point?” – are the ones that catch the assumption baked into step three.
A reviewer who has never said no hasn’t been reviewing. It’s been agreeing. Quietly, consistently, and at scale. And you built a process around it.
The crit where nobody pushes back is a wasted hour. You know it was wasted. The AI review where nobody pushes back feels like due diligence. That’s what makes it dangerous.
If you’re less nervous before design crits than you were a year ago, ask yourself why. If the answer involves AI pre-review, that’s not evidence that the work improved. It might be evidence that the feedback loop got shorter.
If your feedback loop has stopped being uncomfortable, it has stopped being a feedback loop.
