We removed our AI's best feature. Engagement doubled.
When we first built Murror, our AI would automatically surface emotional patterns after every journal entry. "You've mentioned feeling anxious about work 4 times this week." "Your mood tends to dip on Sunday evenings."
Users loved it in demos. They called it "impressive" and "insightful." Our internal metrics looked great -- pattern detection accuracy was 87%.
But something was off. People were journaling less over time, not more. After two weeks, entries got shorter. After a month, many users stopped entirely.
We ran exit interviews. The answer surprised us: people felt watched.
One user said, "I started filtering what I wrote because I knew the AI would analyze it." Another said, "It felt like journaling for a therapist who's always grading you."
The feature we were proudest of was making people less honest with themselves.
So we did something that felt radical -- we made all AI insights opt-in. No automatic summaries. No unsolicited pattern detection. Instead, we added a simple prompt users could tap when they were ready: "Want to see what your entries might be telling you?"
The results:
Journal entry length increased 2.3x
- Daily active journaling went up 41%
- The users who DID tap for insights engaged with them 3x longer
The lesson we keep relearning: in emotional products, trust beats intelligence. The best AI is the one that waits to be invited.
We see this pattern everywhere in our product now. Every time we push information at users, engagement drops. Every time we let them pull it when they're ready, it goes up.
Has anyone else found that making AI less visible actually made it more valuable?
Replies
this tracks with something I've noticed in my own note-taking apps - the moment software starts summarizing me back to myself, I start performing for it instead of using it. curious whether the opt-in prompt itself ever starts to feel like pressure over time, like users start associating tapping it with "time to be evaluated" even though nothing pushed it on them. or does the fact that they chose to tap it change the psychology enough that it doesn't recreate the same watched feeling?
Murror
@galdayan That's such a sharp observation -- the "performing for the app" thing is exactly what we saw too. To your question: so far, the data suggests that user-initiated analysis feels different from auto-triggered analysis. Our 90-day retention for the opt-in group is actually higher than before. We think the key difference is that tapping "show me" feels like curiosity, while auto-summaries felt like being watched. That said, we're keeping an eye on it. If we see the opt-in rate declining over time without a corresponding drop in journaling quality, that could signal exactly the fatigue you're describing.
@monatruong_murror that opt-in rate over time is the metric I'd watch too - it's basically a leading indicator for the fatigue before it shows up in journaling quality. curiosity vs being watched is a really clean way to put words to the difference. good luck keeping it that way as usage scales up
Murror
@galdayan Totally agree -- opt-in rate as a leading indicator for fatigue is something we're now tracking weekly. It's a much earlier signal than journaling quality dropping, so it gives us time to adjust before users silently disengage. And thank you for that framing of "curiosity vs. being watched" -- it's become kind of a north star for how we evaluate new features internally. Appreciate you thinking through this with us!
@monatruong_murror It's not just with AI, but in general as well. A feature can look cool but if it doesnt create the hook, retentioncool,l be poor.
In 2016, we redesigned missingkids.org. I suggested that we add a missing kids' live counter front and center on the landing page that reflects kids missing in that state. It felt emotional initially, leading up to 120% donations. But during usability tests, we saw that it created more anxiety. Hence, we decided to drop it.
Murror
@roopesh_donde That missingkids.org example is a perfect parallel -- emotional impact that initially looks like engagement but actually creates anxiety and drives people away long-term. We saw the exact same dynamic. The 120% donation spike vs. the anxiety it created is such a clear case of short-term metrics lying to you. In our case, people called the AI insights "impressive" but then stopped journaling. The metric that mattered wasn't "do users like this feature" but "does this feature make them use the core product more." Great reminder that this pattern shows up everywhere, not just in AI.