Murror
AI that helps you understand your emotions & connect deeper
Start new thread

We trained our AI to say "I don't know" — and engagement went up.

by•

When we first built Murror's reflection AI, we optimized for insight. Every journal entry got a thoughtful, confident analysis. Pattern recognition, emotional connections, suggestions for growth.

Users were impressed. But something felt off.

We noticed that our most thoughtful journalers -- the ones writing about genuinely complex emotions -- were engaging less over time. They'd write a deep entry, get a polished AI response, and then... nothing. No follow-up. No continued reflection.

We dug in and realized: the AI's confidence was closing conversations instead of opening them.

When someone writes about feeling disconnected from their partner, they don't need an AI that says "it sounds like you're experiencing attachment anxiety rooted in your childhood patterns." They need space to sit with the uncertainty.

So we retrained our approach. Now, when the AI detects something genuinely complex or ambiguous, it says versions of "I'm not sure what this means for you" or "this seems like something that might take time to understand." It asks questions instead of providing answers.

The results surprised us. Users who received "I don't know" responses wrote 40% longer follow-up entries. They returned to the same theme more often over the next week. And in qualitative feedback, they described the experience as "feeling heard" -- even though the AI had literally said less.

The lesson for us was humbling. Our AI was most useful not when it demonstrated understanding, but when it created space for users to develop their own understanding.

In a market where every AI product is racing to be smarter, more insightful, more impressive -- we're finding that the competitive advantage might actually be knowing when to step back.

Has anyone else found that reducing AI confidence actually improved the user experience?

168 views

Add a comment

Replies

Best

This reminds me of good coaching. The most helpful response isn't always the most complete one. Sometimes the right question keeps people thinking longer than the right answer

 Exactly. The coaching parallel is something we think about a lot. A great coach doesn't give you the answer — they ask the question that makes you sit with what you already know but haven't faced yet. We're trying to build that same instinct into Murror's AI. Not smarter responses, but more honest ones.

Very interesting finding. I think many AI products are optimized to give answers, but sometimes users aren't actually looking for answers. They're looking for deep insights, or even just someone/something that helps them think.

In those cases, asking the right question can be much more valuable than giving a confident answer.

 100%. We've started thinking of it as the difference between "answer mode" and "companion mode." Most AI products default to answer mode because it's easier to measure and demo. But for something like emotional reflection, the value isn't in the output — it's in the process. When someone journals about a hard day, they don't need a diagnosis. They need to feel like someone is genuinely sitting with them while they figure it out.

Idk, to say that "They need space to sit with the uncertainty." Sounds like a generalized absolute and that doesn't sit well with me, mainly because as a self reflective user, that's exactly what I'm trying to solve for, so are you reintroducing the problem in a feedback loop?

The phrase sounds punchy, but how are you so sure that's the case? I mention it because personally I don't need more uncertainty, especially if I'm trying to learn about myself, I tend to prefer clear, well defined answers, or even suggestions that help me think deeper to clear uncertainties and move forward. Questions are a great way forward if you don't want to be absolute about it or if you can be transparent about some measure of confidence about analysis.

Sometimes I stop engaging with a tool because it works and has solved my problem well, increasing the chances that I'll use it again because it has built that trust. For something like this I'd worry about using engagement going up as a sole defining metric of improvement, especially if the idea is to fundamentally help people as a measure of success. How do you know it's not creating a dependency loop? And at that point are you building to fundamentally help people or just yourself? Because if so that may be harming your product and users in the long-run.

For example if I notice a model is farming me for engagement I'm less likely to use it or trust it enough to follow up or show vulnerability and that's after doing the work to understand what's happening, even if it gets me to engage a few more times until I notice it. You may get lucky with users who feel comfortable in that loop, but for those looking to really move forward it may be doing them and your platform's trust a disservice.

I'm not saying that's the case, just something to think about here, sometimes in the pursuit of improving an application it can be easy to hinder what already may be working well. It may be that transparently asking users or testers can give you a clearer signal around that.

 Really appreciate you pushing back on this — you're raising the exact tension we wrestle with internally.

You're right that "sit with uncertainty" can sound like a blanket prescription, and it shouldn't be. Some people genuinely need clarity and direction, not more open-endedness. What we've found is that the AI's job isn't to withhold answers — it's to be honest about what it actually knows vs. what it's pattern-matching. When someone writes about a specific conflict, a confident "this is attachment anxiety" can feel insightful but might be completely wrong. Saying "I'm not sure what's driving this, but here's what stands out to me" turned out to be more useful and more honest.

On the engagement concern — you're touching something we think about constantly. We don't optimize for engagement as a success metric. The metric we actually track is whether users report feeling more clarity after a session, and whether they apply something from Murror in a real conversation. If someone stops using the app because they worked through what they needed, that's a win for us.

Your point about transparently asking users is spot on. We've started doing exactly that — short check-ins asking "did this help you think more clearly?" rather than assuming engagement = value. Still early, but the signal is much cleaner than session length or return frequency.

Thanks for the thoughtful challenge. This is the kind of feedback that actually makes the product better.

I wonder if this is less about AI humility and more about human psychology. People rarely grow from receiving neat explanations about their lives. Growth often begins when someone helps them stay curious about a question they cannot yet answer.

 That's a really important reframe. You're right — it's less about what the AI does and more about what it allows the person to do. When we stopped giving tidy explanations, users didn't feel less supported. They actually started writing more, exploring further. The AI stepping back gave them room to stay with their own thinking. Curiosity needs space, and confident answers can accidentally close that space down.

This reminds me of great coaching rather than great technology. The best coaches do not rush to interpret every feeling. They create enough space for people to discover something themselves. Your data seems to suggest AI can do the same.

 That phrase — "create enough space" — is exactly how we think about it. The hardest part of building Murror's AI wasn't making it understand emotions better. It was teaching it to resist the urge to fill every silence with an interpretation. Turns out the space itself is where a lot of the growth happens. Thanks for putting it so clearly.