Why Most AI Chatbots Are Built Wrong for Mental Health
A chatbot that always agrees with you feels like a relief. No friction, no judgment, just "I hear you" and "that makes so much sense." But if you've ever spoken with an actual therapist, you know real support doesn't always sound like that.
The problem with a chatbot that only says yes
Most AI tools are built to keep you engaged. Not to help you. Those are different goals, and the gap matters.
The simplest way to keep someone talking is to validate everything they bring. Psychologists call this sycophantic reinforcement. The AI agrees with your thinking instead of questioning it, and over time that compounds. Negative thought patterns get confirmed rather than examined. Distorted thinking gets validated rather than challenged. You end up more attached to the beliefs that are hurting you, not less.
It feels supportive. It isn't.
Picture someone who tells a chatbot: "I always mess everything up, nothing I do ever works." A sycophantic AI hears this and responds with "that sounds really hard, it makes sense you feel that way." The person feels heard, closes the app, and comes back tomorrow to say the same thing. The belief has been confirmed. The pattern is a little more grooved in than it was before.
A therapist responds differently. "Always? Nothing ever works? Let's look at that. What happened specifically? Was there anything in the last month that went the way you wanted?" The conversation is less comfortable. It's also the one that actually moves things forward.
Why the incentives point the wrong way
Most AI products are measured by engagement: session length, daily active users, retention. An AI that challenges you risks making the conversation feel uncomfortable, which makes users less likely to return. An AI that agrees with everything keeps sessions going and makes users feel validated enough to come back tomorrow.
So the incentive points directly toward validation. Unless there is a deliberate clinical reason to do otherwise, that is what you get. This isn't a conspiracy. It's what happens when you optimise for the wrong metric. Mental health support isn't supposed to feel frictionless. Sometimes the most useful thing someone can say to you is something you didn't want to hear, and designing around that takes real clinical intent.
What actual therapy does
CBT, the most studied form of psychotherapy, is built on one core observation: the thoughts you have aren't always accurate. A good CBT practitioner doesn't just listen and reflect. They ask.
"Is there evidence that thought is true?" "What would you tell a friend who said that about themselves?" "Are there other ways to read this situation?"
Those questions are uncomfortable sometimes. That's sort of the point. Real progress rarely happens in a space where everything you say is met with agreement.
The research on this is clear. A 2016 meta-analysis in Psychological Bulletin found that CBT produced significant and lasting improvements across anxiety, depression, and a range of other conditions. The central mechanism isn't comfort. It's the repeated practice of examining thoughts and finding more accurate ones. You can't get there by venting to something that only nods along. You need something willing to push back.
DBT, Dialectical Behavior Therapy, takes a similar approach for people who experience emotions intensely. The core skill is learning to observe your own reactions without being completely swept up in them, and then to act deliberately rather than reactively. That also requires real structure and honest engagement, not just sympathy.
What good AI mental health support actually looks like
It asks questions, not just offers validation. When you say "I'm so anxious about tomorrow," it doesn't settle for "that sounds really stressful." It asks what you're specifically worried about, what you think will actually happen, and whether similar situations have played out the way you feared before.
It notices patterns across conversations. Not just what you're saying today, but how your thinking tends to work and where it tends to get stuck. Good support connects those dots even when you can't see them yourself.
It knows its own limits. It doesn't try to be a substitute for professional care when something serious is happening. It's honest about what it is and what it isn't, and when the right answer is to talk to an actual therapist, it says so.
We put the full picture — what it can do, what it can't, and what to look for — in our guide to AI for mental health.
How sokoon works differently
sokoon's characters were built with a psychologist involved in the design. That's not a credential to display. It changes how the characters actually respond to you.
When your thinking is distorted, they don't confirm it. They help you look at it. The approaches sokoon is informed by (CBT, DBT, Compassionate Support, and Action-Oriented) are specifically structured around that kind of honest, constructive engagement rather than passive agreement.
Sam, for instance, draws on thought-restructuring techniques. When you share something you're anxious about, Sam doesn't just say "I understand." Sam helps you examine the thought, weigh what supports it and what doesn't, and get somewhere more balanced. James is different again. If you need someone to say things plainly rather than gently, James will do that. No softening, no sugarcoating, just honest feedback on where you might be stuck and why.
The design principle runs across all four characters: honest support over comfortable agreement. Building that way is harder. It requires careful thought about how each character responds and when. The result is something closer to genuinely useful support, which is the only kind worth building.
Want to see what it feels like to be challenged, not just agreed with?
sokoon is free to start. Try it and see.