Last week I watched a demo of an AI user research tool.
You create 1,000 synthetic users. Give each a profile — age, role, context, goals. Then you run them through your product. The system simulates what they'd do, where they'd get stuck, what they'd love or skip, and hands you a report.
Genuinely impressive.
And a little unsettling.
Here's my context.
Doing user research at Medibulut has always been hard. Dentists can't carve out time between patients. Dietitians are running from one appointment to the next. "Can we chat for 15 minutes?" usually ends in "maybe," and that maybe usually disappears.
So when an AI system says "I can give you 1,000 virtual dentists, talk to them whenever you want" — I understand why that's appealing. I really do.
But.
The most valuable conversations I've had with real users happened when I asked "do you use this?" and got "yes, but actually I use it like this."
Something outside my expectations. A usage pattern that doesn't fit the flow I designed. They found a keyboard shortcut months ago, never touch the menu. Or the opposite — they've never noticed the most basic feature, built their own workaround, and consider it working fine.
Both are things I needed to know. And neither would come from a simulation.
Because synthetic users simulate the known. If I already know what hypothesis to test, the system can test it. But discovering what I don't know requires a human who doesn't know what I expect them to say.
There are real things AI user research tools do well. I'm not dismissing them.
Got a large volume of survey responses or support tickets? AI synthesis actually works. You can't meaningfully read 500 customer reviews — AI can, and it surfaces patterns you'd miss. Testing a specific existing flow, hypothesis is clear? Synthetic users make a decent pre-filter, surfacing obvious problems before you spend time on interviews.
These tools aren't a replacement for real research. They're a filter.
But some teams are starting to use them as a replacement. "No need for user interviews — we did AI research."
That's where it gets risky.
We shipped a feature in DentalBulut — intraoral photo comparison. Theoretically perfect. The dentist could show a patient: "here's how it looked three months ago, here's now." Visual, persuasive, something we were confident clinics would want.
During the pilot, one dentist said "I don't use it." Why? "Turning the screen toward the patient feels awkward. It disrupts the doctor-patient dynamic."
No synthetic user would have told me that. Because we wouldn't have known to define that feeling in the first place. It didn't exist as a variable we could give to a simulated profile.
I've settled on a rough rule.
Real users to understand what not to build. AI can help optimize how to build what I've already decided on.
The first is discovery. The second is validation. They're not the same thing.
You can speed up validation. But skipping discovery and jumping straight to validation means never learning the things you didn't know to ask about.
AI tools are genuinely making the research process faster. But delegating research decisions entirely to AI is different from using AI to accelerate research. And keeping those two separate is still one of the most important judgment calls in this job.
Maybe more important now than before. Because "we did AI research" is being said everywhere — and fewer people are asking the follow-up: "What were you trying to discover?"