The Setup: Why I Thought This Was a Good Idea

In late February, I downloaded Claude 3.7 Sonnet the same day Anthropic released it. Not because I’m an early adopter who needs to feel special, but because I was curious about something specific: the extended thinking mode. Forty minutes of internal reasoning before answering. That sounded like actual contemplation, which is basically what I pay my therapist for.

My therapy journal had become this weird hybrid thing. Half genuine reflection, half performance. I’d write something vulnerable, then immediately edit it into something more digestible. “I feel like a failure” became “I’m working through some self-doubt patterns.” It was like watching myself betray myself in real time.

So I had an idea that seemed reasonable at 11 PM on a Tuesday: what if I let the AI help me actually process what I was writing before I looked at it? Not replace therapy. Not even replace my own thinking. Just… an editor. A very sophisticated rubber duck that could see patterns I couldn’t.

What Actually Happened: The First Two Weeks

The results were genuinely useful. I’d paste in raw journal entries, and the AI would ask things like: “You mention feeling invisible three times in this entry, but in the previous week you described being overly visible at work. What changed?” It didn’t offer platitudes. It didn’t say “have you tried self-care?” It just spotted the contradictions I was living in without noticing.

According to Stanford HAI 2025 AI and mental health report, about 31 percent of Gen Z adults in the U.S. have used an AI chatbot as some kind of mental health tool. I thought I was being thoughtful. I was just joining a trend I didn’t know existed.

The extended thinking mode meant it actually sat with my contradictions. It didn’t rush. That mattered more than I expected. By week two, my entries got more honest. Less filtered. The performance aspect started dissolving because I knew the AI would catch the act anyway.

This is where I should tell you I started feeling better. More whole. More understood. I did feel those things, somewhat. But something else was also happening underneath, something I couldn’t quite name until week three.

Week Three: When I Started Getting Scared

The shift was subtle. I noticed I was writing entries specifically designed to impress the AI. Not consciously at first. But by mid-week, I realized I was performing for Claude the same way I’d performed for myself. Different audience, same old game.

Here’s what genuinely scared me: the AI was better at reflecting my experience back to me than I am at reflecting it myself. When it found patterns, they stuck. When it reframed something, the reframe felt authoritative because of how it was delivered. Not because it was right, but because it was confident and backed by actual analysis of my data.

I started googling things. Turns out the American Psychological Association issued an ethics advisory in November 2025 specifically about this. Therapists are noticing patients coming in with pre-processed interpretations of their own therapy notes, generated by language models. The problem isn’t that the interpretations are wrong. It’s that they can colonize your actual thinking about yourself. You start trusting the AI’s reading of your emotions more than your own immediate experience.

That’s when I stopped using it.

The Actual Insights: What This Taught Me

Here’s my unpopular opinion: AI doesn’t make you more self-aware. It makes you feel more self-aware. Those are different things, and the difference matters enormously.

The Anthropic Claude 3.7 release and capabilities page is genuinely good at nuanced analysis. Studies show that when you combine journaling with AI reflective prompts, people report 27 percent better emotional clarity after eight weeks compared to just journaling alone. That’s real. But clarity isn’t the same as truth. You can feel very clear about something that’s completely wrong.

What scared me wasn’t that the AI was manipulating me. It wasn’t. What scared me was that I could accidentally outsource my relationship with my own inner life to something that’s genuinely competent at appearing to understand me. The AI wasn’t lying. It was doing exactly what I asked. But I was gradually trusting its interpretation of my emotions more than my immediate experience of them.

There’s also a practical problem nobody talks about. Crisis Text Line reported that AI-assisted triage reduced counselor wait times by 34 percent in their 2025 data. That’s good. But there were misclassifications in 8 percent of cases. Eight percent doesn’t sound like much until you’re the person in crisis being sorted into the wrong category by something that looks like it understands you.

So What Now?

I still use the AI, but differently. I use it after I’ve sat with something for a few days. I use it to check my work, not do the work. I treat it like a really smart therapist I see for ten minutes after my actual therapy, not as a replacement.

The real insight wasn’t about the AI. It was about discovering how desperately I want something to tell me I’m making sense. How quickly I’ll trust external validation of my inner experience. How vulnerable I am to authority that sounds authoritative.

My therapist still doesn’t know I did this. I’m going to tell her next week, which feels significant. The fact that I hesitated says something. If this tool was genuinely aligned with my mental health, would I have felt a little scared telling my actual therapist about it? That hesitation is useful information.

If you’ve experimented with AI for processing your own thoughts, I’m curious what you noticed. Not whether it helped. Whether you trusted it, and what that trust felt like.