We gave users a button to erase everything their AI knows. Our retention went up.

Three months ago, we added the most terrifying feature in Murror's history: a single button that lets users permanently delete every insight, pattern, and memory their AI companion has built about them. Months of emotional context, journaling patterns, relationship dynamics -- gone in one tap.

Our investors thought we were insane. Our product team debated it for weeks. The data team warned us we'd lose our most valuable asset: the personalization layer that makes Murror's reflections feel meaningful over time.

We shipped it anyway. Here's why.

When we surveyed users who had been with Murror for 3+ months, we kept hearing the same concern phrased different ways: "I've shared things with this app I haven't told anyone. What happens to all of that?" Not in a panicked way -- more like a slow, growing unease. The same feeling you get when you realize someone knows too much about you and you're not sure you consented to it.

For a journaling app built on emotional vulnerability, that unease is existential. If people don't feel safe writing honestly, they start self-censoring. And a journal where you self-censor is worthless.

So we built the reset button. One tap, confirmation screen, everything gone. The AI starts fresh like you just downloaded the app.

What happened next surprised us:

• Only 4% of users actually pressed the button

• - But 67% told us in surveys that knowing it existed made them feel safer

• - Journal entry depth increased 23% in the month after we launched it

• - 30-day retention improved by 11%

The button most people never press became one of our most important trust signals.

We've started thinking of it as the "fire exit principle." People feel safer in a building when they can see the exits, even if they never use them. The mere presence of a way out makes people more willing to go deeper.

This changed our whole approach to AI trust. We stopped asking "how do we make users trust us?" and started asking "what would we need to give users so they don't have to trust us?" Verifiable control over their own data turned out to be more powerful than any privacy policy.

The irony is that by making it trivially easy to leave, we made people want to stay.

For anyone building AI products that handle sensitive data: what does your "fire exit" look like? Do your users know where it is?