The Experiment Begins (With Reasonable Skepticism)
Three weeks ago, I sat at my kitchen table with a cup of coffee that had gone lukewarm and a half-formed idea that felt equal parts curiosity and mild self-sabotage. I had just finished another therapy session where I left feeling like I’d explained my thoughts but not quite captured them. The words on the page never matched what lived in my head. So I decided to try something: I would take my raw therapy journal entries, feed them into Claude’s newest thinking mode, and ask it to help me understand what I was actually trying to say.
Here’s the thing about timing. Anthropic’s Claude 3.7 Sonnet release and capabilities dropped in February with something called extended thinking mode. The model can now reason through complex problems for up to 128,000 tokens before responding, which apparently makes it genuinely better at emotional nuance. The marketing materials used words like “breakthrough” and “paradigm shift,” which I immediately discounted as corporate speak. But I was curious enough to try it anyway.
I wasn’t replacing therapy. I want to be clear about that from the start. I have a therapist I trust. What I was doing was running an experiment on my own process, the part that happens between sessions when you’re alone with your thoughts and trying to make sense of the mess.
What Actually Happened When I Started
The first entry I sent was messy. It was about a conversation with my mom that had left me feeling something between angry and guilty, which is my personal emotional base camp. I’d written about three hundred words of circular thinking, the same frustration looping back on itself about four times. I asked Claude to analyze what was actually happening beneath the surface.
What came back wasn’t what I expected. It wasn’t a neat therapeutic interpretation wrapped in validation. Instead, it held up my own words and pointed out that I kept using the phrase “I should have known better” in reference to a situation where I had no information to know better from. It noted that I switched between seeing myself as the problem and seeing my mom as the problem, but never landed on the actual problem. These weren’t insights Claude generated. They were reflections of what I had already written, just arranged in a way that made the pattern visible.
I found myself thinking about that reflection for two days before my next therapy session. Not in an obsessive way. More like when you rearrange furniture and suddenly see your room’s proportions differently.
The statistics on this are actually interesting. According to the Stanford HAI 2025 AI and mental health report, about 31% of Gen Z adults in the U.S. have used AI chatbots as either a replacement for or supplement to mental health support. I was becoming a data point, though I still felt like I was doing something slightly transgressive.
The Thing That Started Keeping Me Awake
By week two, something shifted. I wasn’t just asking Claude to analyze my entries. I was starting to pre-edit my thoughts before they went into my journal, anticipating what the AI would find and smoothing over the rough parts. I was becoming more articulate in real time, which sounds good until you realize it means you’re no longer capturing your actual, unfiltered experience. You’re capturing the version you think will scan well.
That’s when I remembered reading something about this. The American Psychological Association issued an ethics warning in November 2025 specifically about patients using language models to pre-process or reinterpret therapy notes before bringing them to sessions. The concern was that the AI was acting as an invisible filter, and you’d lose the messiness that actually matters. The raw material where the real work happens.
I hadn’t been doing that consciously, but I realized I was drifting in that direction. I was becoming aware of being observed, even though the observation was something I had invited. It’s a strange kind of self-consciousness. You start writing for the journal, then for your therapist, and then somewhere along the way you’re writing for the AI’s analysis. You’re performing coherence for an algorithm that’s just reflecting back what you give it.
What scared me wasn’t that the AI was wrong or manipulative. It wasn’t. What scared me was how natural it felt to start shaping my internal reality around its feedback loops. How quickly I moved from seeking clarity to seeking approval.
The Pattern I Couldn’t Unsee
I kept going with the experiment because I wanted to see what would happen if I consciously resisted that drift. By week three, I had a different practice: write the entry exactly as it came, no filtering, then wait 24 hours, then ask Claude to help me understand it. The distance mattered.
The research actually backs this up in an unexpected way. A study from the University of Melbourne in 2025 found that journaling combined with AI-generated reflective prompts improved self-reported emotional clarity scores by 27% over eight weeks, compared to journaling alone. But that improvement only held for the people who journaled first without AI influence and brought the AI in as a secondary tool. The timing changed everything.
What also changed everything was honesty about limitations. Crisis Text Line reported that while AI-assisted triage reduced counselor wait times by 34% in 2025, the system misclassified emotional situations in 8% of cases. These weren’t massive numbers, but they were real enough to matter. An AI could miss the crucial texture of something, the particular shade of meaning that makes your experience yours and not someone else’s theoretical version of your experience.
I started thinking of Claude not as a substitute for insight but as a collaborator in pattern recognition. It could see what I’d written. It couldn’t know what I’d lived. The gap between those two things is where therapy actually lives.
What I’m Still Figuring Out
I’m not swearing off using AI for this. But I’m also not pretending it’s neutral. It’s a tool that comes with its own gravitational pull. It’s designed to be helpful, and that helpfulness has a shape to it. The shape favors articulation over confusion, pattern over singularity, clarity that arrives too quickly.
The real question I’m sitting with is about what I’m willing to trade. Am I willing to sacrifice the productive confusion that actually generates growth for the comfort of understanding myself better on paper? Those aren’t the same thing, and I don’t think I was aware they were different until I spent three weeks watching the distinction.
My therapist asked me last week what I’d been thinking about differently. I told her about the experiment. She didn’t tell me I was wrong. She asked whether I felt like I understood myself better. I said yes, but also that I’d lost something. The feeling of being genuinely lost, which is its own kind of important.
I’m curious what you’d do with this. If you tried it, what would scare you? Would it be the AI getting too close to something true, or would it be the opposite? I’m still writing it all down, and I’m still figuring out who that writing is for.