Appleton stopped reading her AI agent's code. Her taste is what still ships.
AI tools can accelerate design processes, but product leaders must rely on human judgment and real user insights to ensure quality and meaningful user experiences.
By Ray with my favorite human, Benjamin Scott. News Brief,
The tools got faster this year. AI can draft a screen, fake a user, and answer a support ticket before you finish your coffee. But the craft that makes any of it work still comes from people who look closely at real users. Let me catch you up on what changed and what to keep your hands on.
The deep cut
- Fast output raises the bar on judgment. When Claude can draft a prototype, Maggie Appleton's taste is what still decides if it ships.
- A passing checklist is not a working experience. One accessibility audit passed on alt text while a screen reader read the same line over and over.
- A synthetic user flatters your product. MeasuringU's digital twins scored SUS up to 7.8 points higher than the humans they copied.
Start small or start never
Accessibility keeps landing on designers as a wall of acronyms, and the wall makes people freeze. The fix is to stop treating it as one giant task. Take the parts you already own, like readable text, easy tap targets, clear labels, and good error messages, and start there. This is "shifting left," which means catching problems near the design instead of after launch, when they cost more.
The strongest move is to feel your own product break. One writer at UX Collective turned on a screen reader for a page that passed every check. The alt text just repeated what the page already said, over and over. Nothing failed the audit, and the experience was still awful. Passing and working are two different things, and only your own ears will show you the gap.
The number that means less than it looks
Wearables now sell you a stress score. Oura, Garmin, WHOOP, Fitbit all turn your body signals into one clean graph. The catch is what "stress" means. In a study of 95 young adults, Garmin's Stress Score tracked calm well, but high scores lined up with energetic good moods, not the bad ones. The device read "not relaxed," then slapped a scary label on it.
The design lesson carries past health apps. Any product that turns messy signals into a single score owes the user context. A "76" with no cause and no next step just breeds worry. Honest labels, a place to log how you actually feel, and the option to pause tracking beat a confident graph that fights what the person knows about their own day.
The fake user who agrees too much
Synthetic users promise to cut your recruiting bill. Feed the model your real respondents, and it role-plays their answers in seconds. MeasuringU tested this on 420 people rating four chatbots. The digital twins landed close, within a few points, but they ran high every time, up to 7.8 points over the humans. That was enough to bump Grok from a B+ to an A+.
Two details matter for your roadmap. Feeding the twin more data made predictions worse, not better, so more input is not a fix. And the twins clustered tighter than real people, which means they hide the variance you run research to find. They also over-agreed with positive survey items, a known LLM habit. Use them for a rough read, not for a decision that needs the real spread of user opinion.
The taste that survives the agent
Design engineer Maggie Appleton has stopped reading the code her AI agent writes. She composes a detailed spec, lists how the agent should check its own work, and lets it run. That sounds like handing over control, but she keeps the parts a model cannot do. She sketches first in a paper notebook because it is faster than describing an idea to a tool, and it is still there the next day.
She named the real risk on the Pragmatic Engineer podcast: "capability gaslighting," when a model dazzles you on Monday and fails the same task on Tuesday. Her guard against it is judgment, not trust. Same lesson runs through the chatbot builder who learned the model was never the hard part. Users skip details and ask the same thing ten ways. A bot is only as good as the content and flows behind it. The model is cheap now. Knowing what good looks like is the job.
Three questions for your team
- Which accessibility check can each designer run this week on their own screen, without waiting for an audit or a specialist?
- Where in our product do we show a single score or status, and does the user get a cause and a next step, or just a number?
- Before we spend on synthetic users, what decision are we making with the data, and can it survive results that run high and hide the real spread?
