The Experiment Started Small

Three weeks ago, I did something that felt both ridiculous and inevitable. I took my actual therapy journal—the one where I write the messiest, most unfiltered version of what’s happening in my head—and started running it through Claude 3.7 before my weekly sessions. Not to replace my therapist. Not even to prepare for sessions, really. I was curious whether an AI with extended thinking capabilities could help me see patterns I was missing in my own emotional debris.

The timing felt right. Anthropic had just released Claude with their new extended thinking mode, which means the AI can reason through complex problems for up to 128,000 tokens before responding. In plain language, that’s a lot of computational thinking. I’d read about it and thought: what if I actually tested this on something that mattered to me? What if I let it engage with real vulnerability, not hypothetical scenarios?

I wasn’t alone in this impulse. A Stanford report on AI and mental health found that roughly one in three young adults in the U.S. have turned to AI chatbots as either a primary tool or supplement for mental health support. The number surprised me less than it probably should have. We’re living in a moment where that option exists, and pretending we’re not using it feels less honest than actually examining what happens when we do.

What Actually Changed in the Editing

The first week, I submitted a three-paragraph entry about anxiety spiraling during a work conversation. The AI didn’t edit my grammar—my grammar was fine. What it did was ask me clarifying questions embedded in a reframed version of my own words. It held up my entry like a mirror at a slightly different angle.

Where I’d written “I felt stupid and everyone could see it,” the response highlighted that I was conflating internal judgment with external perception. It wasn’t correcting me. It was slowing down my narrative long enough to notice the slippage. Over three weeks, this pattern continued. The AI would identify moments where I was catastrophizing, where I was generalizing from one incident to a permanent character trait, where I was skipping the actual feeling to land directly on the story I told about the feeling.

A University of Melbourne study found that people who combined journaling with AI reflective prompts showed a 27 percent improvement in emotional clarity over eight weeks compared to people who just journaled alone. I can feel why that number exists. The AI wasn’t making me feel better. It was making me feel more precise about what I was actually feeling. There’s a difference.

By week two, I started noticing something else. I was writing differently knowing the AI would analyze it. Slightly more articulate, slightly less raw. The self-awareness was creeping in. I was journaling for an audience of one AI instead of journaling purely for myself. That shift, small as it was, bothered me in ways I didn’t expect.

Where the Fear Actually Started

The American Psychological Association issued an ethics advisory last November warning therapists about clients who arrive at sessions having pre-processed their own material through large language models. The concern isn’t straightforward. It’s not that AI will give bad advice. It’s that you might arrive at therapy having already narrated your own story through a particular interpretive lens, and your therapist never gets to meet the raw version.

This is what genuinely scared me. By my third week, I was noticing that I brought Claude’s language into my session. When my therapist asked what happened, I found myself reaching for the AI’s reframed version instead of the original chaos. The AI had helped me organize my thoughts, but in doing so, it had also sanitized them. It had made them more coherent, more digestible. Less true.

I’m not saying the AI was lying. But it was applying order to something that was supposed to stay messy for a minute. Therapy works partly because you sit with the mess long enough for someone else to sit with you in it. If I’d already tidied it up, renamed it, found the silver lining—what exactly was I asking my therapist to help me with?

There’s also the risk that nobody wants to talk about directly. Crisis Text Line reported that when they integrated AI-assisted systems for triage, they reduced wait times by 34 percent. That sounds excellent. But in 8 percent of cases, the AI misclassified the emotional severity of what someone was experiencing. Eight percent might sound small until you imagine being the person in that eight percent, having your crisis sorted into the wrong priority tier.

What I Actually Learned by Stopping

Here’s what I did after week three. I stopped submitting entries to the AI. I went back to just journaling. Raw, unsorted, grammatically questionable journaling. I wanted to see what the difference felt like.

The difference was that I had feelings I couldn’t quite articulate. I had contradictions that didn’t resolve neatly. I had days where I wrote “I don’t know what I feel” and stopped there instead of pushing for resolution. That confusion, it turns out, was important. It’s the texture of actual processing. The AI, in helping me find clarity, had been removing the necessary discomfort.

I’m not saying Claude 3.7 or any tool like it is bad for mental health exploration. The research suggests it helps many people. What I’m saying is that the help it provides comes with a trade-off I hadn’t fully considered. You trade rawness for coherence. You trade process for polish. And if you’re going to do therapy, you might actually need the process more than the polish.

The honest version is this: the AI was really good at what it was designed to do. Articulate, empathetic, consistent, and it genuinely did help me notice patterns. But I realized I didn’t need those things before therapy. I needed them after—as a way to integrate what I’d learned with my therapist, not as a replacement for the initial confusion.

The Bigger Question I’m Still Sitting With

What scared me most wasn’t that the AI was harmful. It was that it was helpful in exactly the way that could make me dependent on it. If I kept using it, I’d probably feel better understood more quickly. I’d arrive at sessions more articulate, more self-aware, less messy. And I’d be outsourcing the most important part of the work—the part where I sit with something true about myself and don’t immediately make it palatable.

I’m not declaring AI off-limits for anyone doing therapeutic work. The research shows real people finding real value in it. But you should know what you’re trading. You’re trading the time it takes to be confused for the comfort of being understood. Sometimes that’s a good trade. Sometimes it’s not.

What I want to know from you is whether you’ve experimented with this. Have you used AI tools in your own therapeutic process, journaling, or emotional work? Have you noticed the shift in how you show up when you know something is analyzing what you write? I’m still figuring this out, and I’m genuinely curious whether my experience resonates or whether I’m overthinking a tool that could just be useful. Come tell me in the comments or send me what you’ve learned. This is unfinished for me, and I think it should be.