The Experiment That Seemed Harmless
I’ve kept a therapy journal for five years. It’s messy, repetitive, sometimes illegible. My therapist says that’s the point. So when Claude 3.7 Sonnet dropped in February with this new extended thinking mode, the one that can supposedly reason through complex emotional nuance for up to 128,000 tokens before even responding, I thought: what if I had an AI editor first? Not replacing therapy. Just polishing my thoughts before I brought them to my actual sessions.
Three weeks seemed like a reasonable test run. I’d write my entries the same way I always did. Then I’d paste them into Claude and ask it to help me organize my thoughts, spot patterns, maybe clean up the circular thinking. It felt productive. Clinical. Safe.
It was neither of those things.
The Seduction of Coherence
Here’s what happened: Claude made my journal entries sound like something I’d actually want to read. It didn’t change the core content, but it smoothed the edges. Consolidated repetitive spirals into neat bullet points. Reframed my anxieties using the kind of psychological vocabulary that makes you feel understood and slightly less broken.
By week two, I noticed I was writing my original entries differently. I was pre-editing for the AI. My thoughts became less tangled because I was already anticipating how they’d be organized. Less raw. The messiness that my therapist had spent years teaching me was actually important started disappearing on its own.
And here’s the part that genuinely scared me: it felt better. That organized version of my mental state was easier to sit with. The AI wasn’t making things up or gaslighting me. But it was making my internal experience more palatable. Digestible. I realized I’d been using AI to create distance from exactly the kind of unrefined emotional data that actually matters in therapy.
A 2025 study from the University of Melbourne found that journaling combined with AI reflective prompts actually did improve emotional clarity scores by 27% over eight weeks compared to journaling alone. But clarity and honesty aren’t always the same thing. Clarity can be a form of self-deception when it comes too early.
What I Learned About AI and Emotional Labor
The thing about extended thinking mode is that Claude could genuinely sit with my contradictions. It could hold multiple conflicting feelings at once without immediately resolving them. That’s actually valuable. But the more sophisticated the AI got at reflecting my emotional complexity back to me, the less I felt I needed to bring that complexity to my therapist.
According to the Stanford HAI 2025 AI and mental health report, about 31% of Gen Z adults are already using AI chatbots as either a replacement for or supplement to actual mental health support. I became part of that statistic almost by accident.
The American Psychological Association started warning practitioners in November 2025 about exactly this problem. Patients were coming to therapy sessions with AI-processed versions of their own experiences. Therapists couldn’t distinguish between what the patient actually felt and what had been smoothed and reframed by a language model. The therapeutic relationship depends on encountering the raw material. Pre-processing it changes everything.
The Honest Reckoning
I showed my therapist what I’d been doing on week three. She didn’t tell me to stop. But she asked me why I felt like my thoughts needed permission from an AI before they were worth bringing to therapy.
That question stayed with me. I think part of it was control. Therapy involves showing someone the chaos inside your head and trusting them to help you make sense of it. That’s terrifying. An AI lets you curate first. You maintain some authority over your own narrative. You’re not as vulnerable.
But vulnerability is the whole point. My therapist’s job isn’t to tell me I’m making sense. It’s to sit in the confusion with me until something shifts. An AI that makes sense too quickly is actually working against that process.
There’s also something unsettling about the Crisis Text Line’s 2025 data. AI-assisted triage cut response wait times by 34 percent, which matters when someone is in crisis. But emotional misclassification happened in 8 percent of cases. Eight percent sounds small until you’re the person being misunderstood in a moment when you need to be understood most.
What Actually Worked
I didn’t stop using Claude entirely. But I changed how. Now I use it after therapy, not before. I ask it to help me understand what my therapist and I discussed. To find the through-lines across multiple sessions. That’s different. That’s synthesis, not filtering.
The extended thinking capability is genuinely useful for this. It can hold weeks of context and spot patterns I’d take much longer to notice alone. But it’s working in service of something that already happened in therapy, not replacing the human part.
What scared me wasn’t that AI is bad at emotional understanding. It’s that it’s getting good enough to be seductive. Good enough that you can convince yourself you’re doing the work when you’re actually just creating a more comfortable distance from it. Good enough that the improvement in clarity feels like progress even when it’s avoidance.
The real question isn’t whether AI can help with mental health. It clearly can, in specific contexts, with real limitations. The question is whether we’re going to be honest about those limitations with ourselves. Whether we’re going to notice when convenience starts looking like wisdom.
I’m curious what’s happened to you if you’ve tried something similar. Not to convince you of my conclusion, but because I might be wrong. I’ve been wrong about technology before. Tell me about it if you’re willing. That part still requires a human.
