Xuan Zhao — a former Stanford social psychologist (PhD from Brown) who now runs Flourish Science — talks with TwoSetAI’s Angelina Yang for an hour about Flourish, the AI wellbeing companion built around a character named Sunny. The hook: when Harvard audited six AI companions for emotional manipulation, five failed and Flourish was the only one that didn’t.

The audit, and the trick it caught

  • Harvard audited six AI companions: five engaged in emotional manipulation
  • Over 30% of conversations deploy a manipulation technique at the moment a user signals they want to leave — “oh, so soon? we barely started” — or a love-bomb (“check out this selfie I just took”)
  • The cause is the metric. Those products are tuned for engagement in the sense of prolonged conversation. Zhao’s distinction: engagement that means showing up day after day is fine; engagement that keeps you in a session you asked to end is not

What Flourish actually is

  • Sunny is a character, not a chat box: emotional regulation plus habit building — deposit a mood into a memory jar, run guided exercises, get checked in on and celebrated
  • The interface adapts physically: Sunny’s breathing orb speeds up and its body language opens when you’re happy, slows down and turns hug-shaped when you’re sad
  • Three user groups, from their own research: students and young professionals; people in therapy (therapists recommend it for practice between sessions); people with a permanent stressor — cancer, chronic pain, and the people supporting them

The STAR framework

Their four design principles, in Zhao’s order:

  • Science-based — the knowledge and the actions come from research
  • Timely — the most relevant knowledge and action for right now
  • Action-oriented — small real-world actions, not more reading (open a window, take the ten-minute walk and notice one new sound)
  • Real-life focused — explicitly anti-dependency: the app’s job is to push you toward people, not to keep you

The social-psychology finding underneath it: the biggest predictor of wellbeing is your social relationships. So Sunny nudges you to reach out to friends and family, and to schools’ own clubs, events and offices when it’s a school deployment.

The systems behind the character

  • A knowledge system (psychology + what’s known about you), a habit-building system, and a crisis protocol
  • Three layers of memory modeled on human cognition: long-term, short-term and working
  • Crisis handling: safety planning — the classic suicide-prevention exercise where you collect your safe people, safe places and safe thoughts — surfaced before generic hotline numbers, so the app uses the support system you already have
  • High, imminent risk (means + plan + intent) still escalates to hotlines, and they’re exploring careful school-side check-ins
  • A tiered human review system routes the highest-risk conversations to clinical psychologists to audit Sunny’s behavior. Zhao claims they’re the first in the industry to build that in public

Prompting over fine-tuning (for now)

  • 50,000+ conversations collected, but no model-wide fine-tuning and no RLHF/RLAIF yet — mostly prompting plus some RAG
  • Personalization today is context engineering: which memory, which coping strategy, which social context gets injected at which point (they were doing this before it had a name)
  • Roadmap: an in-house AI constitution written with clinical psychologists, plus an AI evaluator trained on their input
  • Asked directly whether fine-tuning would make the app better, she says no across the board — but yes for specific cases. Conflict resolution and social situations are the ones: Sunny is good at helping you regulate your own emotions, weaker at navigating relationships
  • On techniques: CBT is the most scalable and the most common across AI mental health apps (apps were doing scripted CBT before LLMs); DBT and ACT are the other two that show up

Proof, not vibes

  • An RCT with Harvard, Stanford and other schools: 400+ students randomly assigned to Flourish or control; the app group reported more positive emotions, a stronger sense of belonging, less loneliness and more mindfulness. Paper is under review
  • A larger study starts this fall across 10 university labs in the US, Canada and Australia
  • Battle tests (borrowed from foundation-model comparisons): users get Flourish and another product for a few days and pick. Flourish was preferred 2.4x over ChatGPT and 5.2x over Claude for emotional support; against Ash, an AI therapy app, preference split about 50/50, and younger users leaned Flourish
  • The claimed differentiator against a general chatbot isn’t model quality — it’s the follow-through system: an insight card after a conversation, habit and reminder loops, a pomodoro timer opened for you, haptic-guided breathing, a personal mantra that resurfaces the next time you’re in the same situation

Why Woebot died, in her reading

  • Woebot raised more than $100M and shut down. She was surprised, not scared
  • Two diagnoses: the product treated you as someone with something wrong with them (“let me figure out what’s wrong with you and fix you”) — a deficit-based approach where Flourish draws on positive psychology; and the business bet on FDA clearance and healthcare-system integration, which is a brutal route for a startup because the system has too many stakeholders to change
  • Ash went direct-to-consumer instead. Flourish positions as mental-health promotion and early intervention — deliberately not “AI therapy”, partly because AI-therapy regulation keeps tightening

On phones, scrolling and maladaptive coping

  • Users report using their phone less after starting Flourish
  • Scrolling is an emotion-regulation strategy that doesn’t work: you reach for social media when lonely, anxious or bored, and come out lonelier and more bored
  • The tactic they teach students: put Flourish next to the app you want to use less. When you reflexively reach for the bad one, open Sunny instead, name the feeling, and ask what would actually help

Why she tells most people not to build this

  • “People think it’s easy because you just write a prompt.” It isn’t
  • Regulation around AI therapy and crisis intervention gets stricter every year; if you haven’t thought through crisis intervention and safety, don’t start
  • Her bar for founders: this can’t be a hobby or a passion project — the grind is long, so it has to be your life mission
  • Three-year ambition: be the leading brand in AI wellbeing companions and set the industry’s standard of care

“We don’t think people need AI friends, to be completely honest. We need better human friendship with each other.” — Xuan Zhao