The Experiment I Thought Would Be Harmless
Three weeks ago, I started doing something I didn’t tell my therapist about. Every evening, I would open my therapy journal—the one I’ve kept for two years, full of the messy, unfiltered thoughts I’m supposed to bring to sessions—and I’d paste entries into Claude 3.7 Sonnet with a simple request: “Help me understand what I’m really feeling here.”
It felt practical. Smart, even. Like I was optimizing my mental health instead of just sitting with discomfort. I’d heard that Anthropic Claude 3.7 release and capabilities included extended thinking mode that could work through complex emotional reasoning in ways previous versions couldn’t. The AI could spend up to 128,000 tokens reasoning through what I’d written before responding. That sounded like exactly what I needed at 11 p.m., spiraling about something from the day.
I’m telling you about this because I’ve changed my mind about it. Not gradually. Suddenly. And I want to be honest about why.
What Actually Happened When I Started Using It
The first two weeks felt good. Genuinely clarifying. I’d write something raw—frustration about a conversation with my partner, anxiety about work, old shame cycling back—and Claude would offer structured reflections. It would identify patterns. It would reframe things. By the time I read the AI’s response, I felt like I’d already processed the experience. I felt understood. I felt lighter.
Then I noticed something strange: my journal entries started changing. Not what I wrote initially, but how I was writing them. I was unconsciously pre-editing for the AI. Making things more coherent. Less messy. Anticipating what would be easier for it to analyze. I wasn’t writing what I actually felt anymore. I was writing what I thought would make sense to an AI that was supposed to help me understand what I actually felt.
That’s the moment I should have stopped. I didn’t.
Instead, I kept going for another week, which is where things got genuinely unsettling. I found myself bringing Claude’s interpretations into my actual therapy sessions. Not as raw material. As conclusions. My therapist would ask me what I noticed, and I’d say things like “Well, Claude pointed out that this is actually a control issue” rather than sitting in the uncertainty of what I was experiencing. I was reporting an AI’s processed version of my emotions back to my human therapist as if it were my own insight.
Why This Actually Scared Me
I want to be clear about something: I don’t think AI is evil or that technology is ruining therapy. That’s not what scared me. What scared me was how easily I let someone else’s analysis—even an AI’s—become my self-understanding.
The American Psychological Association issued an ethics advisory in November 2025 flagging exactly this problem. Patients were using language models to pre-process and reinterpret their therapy notes before sessions, essentially creating an intermediary between their raw experience and their therapist’s understanding of it. On paper, that sounds like nothing. In practice, it means the therapist isn’t meeting the actual person in the room anymore. They’re meeting the person-plus-the-AI’s-interpretation of the person.
Here’s what really landed for me: according to Stanford HAI 2025 AI and mental health report, about 31% of Gen Z adults are already using AI chatbots as a substitute for or supplement to mental health support. The research is still emerging, and I understand why people do it. It’s free. It’s immediate. It doesn’t judge you. But what if the problem isn’t what the AI does to us, but what we do to ourselves when we let it mediate our self-knowledge?
I was using Claude as emotional smoothing. Taking the jagged, contradictory, sometimes incoherent experience of being a human and running it through a system designed to find patterns and make meaning. Which sounds helpful until you realize: therapy isn’t actually supposed to feel coherent all the time. Sometimes the whole point is to sit with contradictions. To notice that you believe two opposing things at once. To feel confused and let that confusion teach you something.
What The Research Actually Shows
Here’s where I need to be intellectually honest: some of this works. A 2025 study from the University of Melbourne found that journaling combined with AI reflective prompts improved people’s self-reported emotional clarity by 27% over eight weeks compared to journaling alone. That’s real. The combination helped.
But there’s something important buried in that finding. It helped when people were combining it with journaling, not replacing journaling with AI responses. When I asked Claude to edit my therapy journal entries, I wasn’t supplementing the process. I was outsourcing a crucial part of it. The part where I have to sit with my own confusion long enough to find my own language for it.
I also read the Crisis Text Line’s 2025 report about AI-assisted triage. The system reduced counselor wait times by 34%, which matters in crisis situations. But it also misclassified emotional severity in 8% of cases. That’s the trade-off nobody talks about much. Efficiency gains paired with the statistical inevitability of error. In mental health, 8% of misclassified cases isn’t an abstraction.
What I’m Doing Differently Now
I deleted the Claude conversations. Not because they were harmful or full of bad advice. I deleted them because I needed to reset my relationship with my own thinking.
Now when I’m stuck in my journal, I sit with it longer. I write the same thought three different ways. I notice contradictions instead of smoothing over them. It’s slower. It’s less satisfying. It’s also mine in a way it wasn’t when I was using an AI as an emotional middleman.
I’m not saying AI can’t be useful in mental health contexts. But the moment it becomes a substitute for your own discomfort with your own experience is the moment it stops helping you and starts replacing you. And the tricky part is that it feels good while it’s happening. That’s what scared me most.
Have you ever outsourced something internal? I’d genuinely like to know what you noticed. What felt efficient, and what did you lose in that efficiency? Drop me a line. I think we should probably be having these conversations with other people instead of just processing them alone with an AI anyway.
