When we shipped Mirror Test, we expected the answers to be flattering. We expected a kind of birthday-card energy — sweet, generic, mostly about being "kind" and "funny" and "always there." That isn't what we got.
What we got, across the first thousand or so completed mirrors, was strange and specific and quietly devastating. People do not describe each other the way they describe themselves. They notice different things. They remember different scenes. And when you give someone the cover of anonymity, they will tell their friend something that has been sitting on their chest for months.
The first surprise: nobody talks about traits
The single most common pattern in the answers is that almost no one writes adjectives. We expected "loyal," "smart," "creative." We got, instead, sentences like:
"You are the only person who texts me back in full sentences. I notice it every time. I don't think you know you do it."
Or:
"You leave a room better than you found it. Always. Even when you didn't say much, the temperature is different after you go."
People write scenes, not labels. They reach for the small concrete behavior. The dish washed without being asked. The voice memo on a hard day. The way you wait for the slow walker. We had to redesign half the report around this — the AI was reaching for adjectives the answers never contained.
The second surprise: contradictions are the most common signal
The third question — "what's a contradiction that feels true about this person?" — was the one we worried about. We thought it would be skipped. It is, in fact, the most-answered optional field in the product.
People are eager — almost relieved — to say the thing they've been holding. The pattern looks like this:
- "You're the most generous person I know and you're terrible at receiving anything."
- "You're funniest when you're scared."
- "You give the best advice and you never take any of it."
- "You are deeply private and you tell strangers your real life within ten minutes."
The contradiction question turns out to be the closest thing to truth-serum the product has. It works because it doesn't ask anyone to judge — it asks them to notice. There's no risk in noticing a paradox; you don't have to decide if it's good or bad.
The third surprise: the report is not the point
We thought the report — the AI-synthesized "social mirror" — would be the moment. The product, the artifact, the thing people would screenshot. It is, sometimes. But more often the moment that people describe back to us is reading the raw answers.
People sit with their phone in a coffee shop and read what their friends typed and they cry a little, or they laugh a little, or they sit very still. The synthesis is interesting. The fact that five people independently noticed the same small thing is what hits.
That changed our roadmap. The report stays — it's the unlock, the carrot, the artifact you can share. But everything around it is being rebuilt around the raw answers as the primary object.
The fourth surprise: people don't ask for praise
When you give people three questions about a friend and tell them no one will ever know who answered, they do not write "you're awesome." They write the thing they wish they had said out loud. They write the thing they have been embarrassed to mean. They write the thing they assumed the person already knew, and then, halfway through, they realize the person almost certainly does not know it.
We thought we were building a personality test. We were not. We were building a low-stakes container for the things people have been meaning to say to each other for years.
What we changed
A few concrete things shifted because of the data:
- We rewrote the third question's placeholder text three times. The current version — "Two things that shouldn't both be true — but somehow are" — gets the longest, most careful answers we've ever measured.
- We made the optional questions more inviting. Skip rate dropped from 38% to 11%.
- We added a "make my own mirror" prompt at the end of the answer flow. About one in five answerers now goes on to make one.
- We stopped asking for adjectives anywhere in the synthesis prompt to the model. The reports got better immediately.
What we won't do
We will not show owners who answered. We will not score people. We will not turn this into a personality framework with five letters and a color. The thing that makes the answers honest is the certainty that no one is being judged on the other side. We are protective of that to the point of being annoying about it.
If you want to try it on yourself: make a mirror. If you want to read a real one before you do: here is the example mirror.