What Designing AI Taught Me About Why Psychiatrists Ignore It

COMMENTARY

I remember sitting in on a product meeting, watching a startup demo a risk-flagging dashboard. Clean interface, live electronic health record (EHR) integration, color-coded alerts flagging high-risk patients. It looked like the future of medicine: artificial intelligence (AI) spotting the patients who need care before any of us has a clue.

Then I started my psychiatry residency, and I watched a nurse swipe away that exact type of dashboard without looking at it. Not out of rejection, but out of survival. She had 4 deteriorating patients and a fifth refusing medications. What I thought was the future of psychiatric care had no idea what the ward looked like at hour 10 of a shift.

The Convenient Explanation

The technology industry typically handwaves moments like this with a reductionist narrative: clinicians are resistant to change. But this framing is commercially convenient. When a tool fails, asking what is wrong with the user costs nothing and changes nothing. It protects the product and pathologizes an entire profession rather than interrogating a design process.

Both Sides of the Screen

I have worked on both sides of this failure. Before residency, I designed AI and user experience (UX) solutions for the United Nations and the World Health Organization to reach populations under stress, navigating high-stakes decisions in chaotic environments. I now manage those same conditions as a psychiatry resident.

Moving between these two worlds, I have come to see how closely clinical empathy and UX empathy mirror each other. Both require you to set down your assumptions and sit inside someone else's reality. A psychiatrist who projects what they imagine a patient needs is not doing good psychiatry. A designer who builds for a user they have never observed is not doing good design. The failures are structurally identical.

Diagnosis

The root cause is a design process that treats the clinical environment as a deployment context rather than a source of knowledge. Skipping that step does not save time. It produces tools that shift cognitive burden rather than reduce it, that serve managerial dashboards rather than frontline workflows, and that get swiped away by clinicians who do not have the bandwidth to explain why. The gap between product and practitioner is not a technology gap; it is an empathy gap, and it surfaces the same way across tools and platforms.

Ambient documentation tools like Nuance DAX or Abridge are often held up as psychiatry's most obvious AI win. The time savings are real, but in psychiatry, the note is not a by-product of the encounter; it is an act of clinical interpretation. An AI summary captures words but not their weight. When the patient and the doctor know they are being recorded, something shifts in the room. A designer who had observed even a handful of psychiatric encounters would have recognized that the clinical note in psychiatry is not the same object it is in cardiology.

Suicide risk prediction models integrated into EHR systems face a different version of the same problem. The algorithms are often statistically sound. But when a quarter of an inpatient unit is flagged as high-risk on any given morning, which in an acute ward is not unusual, clinicians stop reading the flags. Alert fatigue is not a clinical failure. It is what happens when a designer optimizes for sensitivity without asking what a nurse at hour 10 of her shift can realistically process.

Chatbot-delivered therapy platforms like Woebot or Wysa show genuine engagement in consumer populations. But a product optimized to retain users is structurally incentivized to be as validating and comforting as possible, while effective therapy is sometimes neither. And when a patient crosses into acute risk, a chatbot has no architecture for that moment.

The Prescription

The path forward is not simply more collaboration. Collaboration without methodological rigor is just more meetings. What clinical AI development needs is the same standard of user research that medicine applies to everything else it adopts. That starts with sustained ethnographic observation: not a site visit, but enough time on the ward to understand how attention actually moves, what a workstation looks like at peak load, and where the invisible workarounds live. It continues with iterative prototyping that treats frontline clinicians as codesigners rather than validators, which means involving them before the interface exists, not after it has already been built.

It also requires redefining what success looks like: adoption rates and engagement metrics measure whether a tool got used, not whether it helped. Cognitive load, workflow disruption, and time-to-clinical-decision need to sit alongside them. This is standard practice in human-centered design for humanitarian contexts. There is no principled reason it cannot be standard here.

What is less obvious is that psychiatry is unusually well-positioned to drive that shift. Psychiatric training builds exactly the skills human-centered design demands: sitting with ambiguity, noticing what is not being said, constructing a model of another person's experience from indirect evidence under time pressure. The developers who recognize that, and build accordingly, will find psychiatrists more useful collaborators than they expected.

Prognosis

Psychiatrists are not slow adopters. We are accurate evaluators. We are waiting for tools built by people who finally sat with the complexity of our ward and clinic before writing a single line of code. And we look forward to working with developers who extend to us the same quality of empathy they expect us to extend to every patient we see.

Dr Lam is a psychiatry resident in Hong Kong. She previously served as a consultant to the World Health Organization and as an AI and Data Science Fellow at UN Global Pulse.