Welcome to the first Close Read.
I've been a designer for a long time, and these days I also build what I design, mostly with AI. Twice a week I'll take one thing from the AI world and look at it the way I look at any product: who is it for, does it work, and what does it do to the people using it? Sometimes that means praise, sometimes criticism. This first issue is about what happens when AI plays your users.
What 120 synthetic users missed
Nielsen Norman Group ran a simple test. In an earlier survey, 604 tech and design professionals had told them what they would call UX work. NN/g then put the same open question to 120 synthetic respondents: AI-generated stand-ins, created on a platform called Synthetic Users.
Both groups reached the same overall conclusion, so the AI got the headline right.
The difference was in the range of answers. The humans used more than 50 distinct terms for their work, from "journey management" to "human-factors engineering". The synthetic answers varied much less. NN/g calls this mode collapse, and it showed up across the whole sample and within each subgroup.
NN/g's conclusion: "Research with synthetic users is still not user research, because no actual users are being studied." Their advice: "Use synthetic respondents to generate hypotheses or check broad patterns; when the tail of responses matters, you need human participants."
Read the study: https://www.nngroup.com/articles/synthetic-users-vs-human-respondents/
My take
I know NN/g well. I've taken two of their week-long courses, and this kind of small, honest test that anyone could repeat is what they do best.
I'm all for speed. I'd rather put something real in front of people next week than research for two months. So I understand the appeal of synthetic users: you get answers in minutes, without recruiting or scheduling anyone.
The average answer is usually the one the team already guessed, though. What makes research worth it is the person who uses a product in a way nobody planned for, the answer nobody on the team would have written. That's the part that shrank in NN/g's test.
So I'd use synthetic users the way I use AI everywhere else, for the legwork: to sharpen the questions and test the study plan, and to get the obvious answers out of the way so the sessions with real people can go deeper. Where a mistake can hurt someone, I wouldn't take the shortcut at all. Quite the opposite, to be honest.
NN/g notes that more teams are turning to synthetic users, which is rather concerning. If those simulations keep landing on the average, the people at the edges drop out of product decisions without anyone noticing, and they are often the people who most need a product to work.
Also worth a look
ai-job-search, job applications with Claude Code and the battle of agents
A job-search setup built on Claude Code is trending on GitHub, with more than 52,000 stars. It evaluates postings, tailors your CV and cover letters, and preps you for interviews. Its README explains where it comes from: "It got me hired." The README is also refreshingly honest about its limits. The LinkedIn search is "for personal use only", because automated access is against LinkedIn's terms of service, and the defenses against prompt injection are "instruction-level, not a sandbox".
If recruiters screen with AI too, we end up with tailored CVs meeting automated filters. What does a recruiter look at when every CV fits perfectly? And who will win this? The job seekers' agents or the recruiters' agents filtering it out again? This system is obviously broken and will have to change quite significantly in the future. Or we don't have jobs anymore, which would be a solution to this problem as well.
Full disclosure: I made a job application kit for Claude myself, Apply Kit, so I have a stake in this. Feel free to use it, if you feel like it: https://github.com/starburst1977/apply-kit
https://github.com/MadsLorentzen/ai-job-search
Tidy, an open-source design canvas
Tidy is a free, open-source alternative to Figma and Paper. It's under the Apache 2.0 license, and you can host it yourself. Coding agents like Claude Code, Codex and OpenCode can edit designs in it through MCP. It's early, but a design tool built for people and agents from the start is an interesting bet.
Emochi, a voice journal with a character
Emochi turns your voice notes into a "living, generative character" that reflects how you feel (this sounds like fun), and shows patterns in your entries over time (this is probably not so much fun). It was built by "a doctor, a designer, and two AI engineers", and it's free on iOS and Android. It's on my list to try. The first thing I'll look for is where the recordings go: the launch page calls it private, but doesn't say where recordings are processed or stored. But I like the idea and I would see potential here to go even further.
https://www.producthunt.com/products/emochi-2
If you have thoughts on synthetic users, I'd like to hear them. Leave a comment here or reply on Bluesky.