London
Paul Couach
Psychologist, London. I study how minds work, and I build with the new ones. Move across my name.
I’m a psychologist. I took a First in Psychology at Goldsmiths and an MSc in Clinical Mental Health Sciences at UCL. Before that came business school in France and a sales floor in Madrid.
I think general-purpose AI arrives this decade,It’s a forecast, and I’ll update it when the evidence moves. and that it’ll change mental health and ordinary work more than anything else I’ll live through. So I’ve rebuilt how I work around it. A small staff of AI agents writes nearly all my code. My job is to point them somewhere worth going, and to check what they bring back.
The questions I care about sit where those two worlds meet. What happens to people who talk to machines all day? And how does a human stay a useful check on a system that’s smarter than them?
Now Looking for research roles where mental health meets AI: human-AI interaction, model behaviour, digital mental health. London, or remote from London.
A living map of human studies on how AI use affects mental health, graded by what each study can show. It holds 306 studies so far. Only 41 are randomised trials, and in two areas, psychosis and self-harm, nobody has run a trial yet.
An interactive 3D workshop manual for the Jaguar XK8.Built for a phone propped on the wing, with greasy hands. Take the car apart, watch the engine run and follow 43 repair jobs with factory torques, in English or French. Every value shows its source, and missing values stay blank rather than guessed.
My homelab and personal operating system: 28 services on Proxmox, from photo backup to the assistants that run my week. About 4,900 commits since April, nearly all written by agents.Agents are optimistic narrators. The table exists because of that. Nothing counts as working until it clears the last row of this table:
| Level | What was checked | What it proves |
|---|---|---|
| Source | Tests pass in the repository | The code hangs together |
| Deploy | The change reached the machine | It’s installed |
| Live | A fresh end-to-end run on the real system | It works |
I packaged the rule for anyone as Homelab Operator, a small CLI that makes every pull request say what it proved and what it didn’t.
When AI helps you check AI, do you catch more errors or just feel surer? A reproducible study design with model-written summaries, seeded mistakes and 4 kinds of assistance, from a bare answer to a full verification checklist. The participants are simulated for now, so it’s a test of the design.
Change the wallpaper and 14 desktop tools re-theme to match it, with animated transitions. Linux-style ricing for a Mac.
Before the agents, I ran studies on people.
Staff morale in inpatient mental health rehabilitation
How staff in NHS and independent-sector rehab units talk about morale, from interviews and focus groups in the ACER study.
Is Acceptance and Commitment Therapy worth the money for older adults?
A systematic review of cost-effectiveness: screening, extraction, appraisal, and pulling ICERs and QALYs out of economic evaluations.
Does nonsense sound deeper when it comes with a face?
An experiment on how video, pictures and audio change the way people judge pseudo-profound statements.
Mimicry, listening and trust
Across 40 face-to-face interviews, mirroring someone and actively listening to them each raised how much they trusted the interviewer.
Making things got cheap. Knowing which things to make is the job now.
Write to me about minds, machines or old Jaguars.Competing interests: I have opinions about old Jaguars.
[email protected] LinkedIn GitHub