AI for friction, humans for meaning. Raw output is never finished work. Verify before trusting. Disclose where it matters. Skepticism welcome when specific.
Standing recap.
The stakes
UX is where AI's biggest lie lives
Confident personas · quotes from people who don't exist · statistics from studies that never happened
Fake research is worse than no research.
"It carries authority without evidence." Slow delivery.
Board-worthy
The professional split
AI = structure + pessimism Research = evidence + empathy
This is the week. Stays on the board all day.
The week in one line
AI can imagine your users. Only research can know them.
Label which one you're holding.
The label discipline starts now.
Today
Today's promise
A five-prompt UX planning kit that runs on any idea in 45 minutes — and the professional answer to "how do you use AI without faking research?"
That interview question is real and current.
Deliverable
You will leave with
One task flow (with unhappy paths) One low-fi wireframe (grey, real microcopy) A critique checklist for your design
Show the prepared trio briefly.
Today's map
Today's map
Synthetic users vs. research + the audit — 40 min
Edge cases + accessibility specificity — 15 min
Demo: the plant app, five prompts live — 35 min
Lab: Edge Case Hunt — 70 min
Share: trust-damage rankings — 40 min
Breaks protected.
Part 1
Synthetic users vs. real research
~0:10.
The central metaphor
Personas from a model are
Hypotheses wearing name tags.
Useful as hypotheses. Poison as findings.
Returns in the demo.
Non-negotiable
The three rules
Label everything:HYPOTHESIS — needs real research
Every claim gets a test question
Never present synthetic users as research — that's fabrication
Rule 3, flatly: "It's the kind that gets discovered, and it ends careers."
Live debate
The industry is fighting about this right now
"Synthetic user research" vendors vs. research professionals.
The position that survives scrutiny: hypotheses → then verification.
NN/g's skeptical video linked on week page. "Be able to argue their side."
Pre-empted
Why bother with personas at all?
Because articulated assumptions beat invisible ones.
The hypothesis forces you to say what you believe — the only way to find out you're wrong cheaply.
Answer to the inevitable question.
Part 1, continued
The assumption audit
The cheapest UX win
~0:25.
Ten minutes
Ten minutes that reorder a project
"List every assumption hiding in this idea." Then sort:
safe to assume · must research · dangerous to guess
AI at its best — pessimistic structure, no facts required.
Try this now
2 minutes
Audit "[your most annoying app] redesign"
What's it assuming about you that's wrong?
One "dangerous to guess" each — shout them out.
Collect three aloud. Fast energy beat.
Flows
The happy path is the easy 20%
AI drafts a clean flow in seconds.
The value move is the next prompt:
"Now give me the unhappy paths."
Bad photo, no signal, ambiguous result, abandoned mid-task.
Why it works
Models are good at failure
They've read every support forum on earth.
Most of real UX is designing the recovery.
Genuinely true and genuinely useful — pessimism on tap.
Wireframes
Ugly on purpose
Grey boxes. Real structure. Real microcopy.
Polish hides thinking — yours and the model's.
"A beautiful screen with a broken flow gets approved, then fails."
The quiet deliverable
Microcopy is a deliverable
Button labels. Error messages. Empty states.
Where products feel human or hostile.
AI drafts; you de-robot; ⚠ marks your overrules.
The ⚠ overrule mark = this week's authorship evidence.
Break
Break — 10 minutes
Stage mermaid.live + WCAG quick ref tabs.
Part 2
Edge cases + accessibility
Pessimism as a service
~1:00.
The hunt
The Edge Case Hunt
After any flow:
"8 edge cases, ranked by how badly each damages trust."
Trust-damage ranking = your prioritization tool.
Ranking by trust beats ranking by likelihood — felt severity drives users away.
Accessibility
Specific, then verified
✗ "Ensure good accessibility" ✓ "Photo capture with low vision · colour-only status · touch targets in gloves · screen-reader order"
Then verify against WCAG — the source, not the paraphrase.
"Models paraphrase guidelines loosely. The quick ref has the numbers."
The chain
Plan → prototype
flow → wireframe text → boxes → clickable path
AI accelerates every document before the canvas. The canvas is still judgment.
Figma AI features churn quarterly — the planning chain works regardless.
Demo
"An app for plant people"
Five prompts, live
~1:15. Prompts from DEMOS file, in order.
Demo · prompt 1
The audit
Every assumption, sorted
safe / must research / dangerous to guess
Star the dangerous ones aloud.
Demo · prompt 2
Personas
Two hypotheses + their test questions
"These people do not exist. They're hypotheses wearing name tags."
If the room nods along with the personas, catch them: "You believed them. That's exactly the problem."
Demo · prompt 3
Flow
Diagnose-my-plant + 3 unhappy paths
bad photo · no signal at the garden centre · ambiguous result