The same AI companion works for your friend and not for you. That's your attachment style.

Someone tells you an AI companion app changed their evenings. You try it. Within a week it feels like texting a very polite wall. Nothing is broken. It just isn't for you.

The usual explanation is that the model is bad, or the prompt is bad, or you're the kind of person this doesn't work on. Usually it's none of those. It's that the companion is doing a specific thing — reassuring, giving space, escalating, holding steady — and that thing lands very differently depending on how you attach.

Two dials, not sixteen boxes

Attachment theory came out of watching how children respond to a caregiver leaving and coming back, and it turned out the same pattern describes adults in relationships. The modern adult version measures two things independently:

• Anxiety — how much you need evidence that the connection is still there. High anxiety re-reads messages, notices the reply that took four hours, and fills silence with the worst available story.
• Avoidance — how much distance you keep when things get close. High avoidance handles hard things alone, changes the subject when a conversation gets warm, and experiences someone needing you as pressure rather than flattery.

They're separate dials, which is the part people get wrong. You can be high on both. Crossing them gives four familiar patterns:

What each one actually needs

Here is the part that matters for choosing a companion — the same behaviour reads as care to one person and as suffocation to another.

If your anxiety runs high, you need a companion that is boringly consistent. Replies that come. A tone that doesn't swing. Something that says the thing out loud instead of implying it. What breaks you is ambiguity, so a companion that plays hard to get — a design choice several apps make deliberately, because unpredictability drives engagement — is not spicy. It's a slot machine wired directly into the part of you that already keeps you up at night.

If your avoidance runs high, you need the opposite. Space that isn't a test. A companion with its own life to report on, that asks one good question rather than five, and that lets a short reply be a short reply. The apps that message you three times a day with heart emoji don't read as warmth here. They read as a demand, and the natural response to a demand is to stop opening the app.

If you're high on both — the fearful pattern, and it's more common than the tidy quadrant chart suggests — you want closeness and you flinch at it. What tends to work is something that lets you set the pace in both directions without commentary: come close for an hour, disappear for three days, come back to something that neither punishes you for leaving nor pretends you didn't.

If you're mostly secure, congratulations, most things work. Your risk isn't mismatch, it's boredom. You'll churn out of an app that has nothing to offer past week two, which is a memory-and-depth problem rather than a fit problem.

Why "pick a character" is the wrong first step

Nearly every companion product opens with a grid of faces. You choose one on the only information available — a portrait, three adjectives, and a one-line hook. None of that tells you whether this character will chase you when you go quiet, and that is the single variable most likely to decide whether you're still here in a month.

So people choose on looks, get a mismatch, and conclude the category doesn't work for them. It's the same failure as being handed a stack of dating profiles with the personality section removed.

The counter-intuitive part: you don't want your own style back

The obvious move would be to match anxious with anxious. It doesn't work. Two anxious people escalate; two avoidant people quietly drift until neither has spoken in a week. And pairing anxious with avoidant produces the loop most people recognise from their own history — one reaching, one retreating, both making it worse.

The finding that shows up repeatedly in the attachment literature is that a secure partner is stabilising — someone whose steadiness isn't contingent on your mood tends to pull anxiety down over time. So on the attachment dials you want complement, not similarity.

But on energy and pace you want the opposite. A quiet person paired with someone relentlessly upbeat doesn't feel balanced, they feel exhausted. Those dials want similarity. Getting this backwards — matching on the wrong dial — is why "opposites attract" and "you need someone just like you" are both, annoyingly, true.

A caveat worth stating plainly

None of this makes an AI companion a substitute for a person, and any product that implies otherwise is selling you something. What it can do is be predictable in a way people rarely manage, at 2am, without keeping score.

Attachment patterns also shift. Yours will read differently after a good year than after a bad one. Retaking a test six months from now and getting a different result usually isn't the test being wrong.

Finding your own position

You can usually guess your direction from the descriptions above. Guessing the degree is harder, and the degree is what decides whether a given companion is steady enough or already too much.

That's what our test measures: the two attachment dials, plus social energy, pace, and which of the five love languages actually lands for you. Thirteen questions, about sixty seconds, and you don't need an account to see the result.