It's 1 a.m. and something happened today that you'd like to tell someone. Not a crisis. Just the kind of thing that sits badly if it stays inside your head. The people who'd understand are asleep, or you haven't talked to them in eight months, or you don't want to be the friend who only surfaces when something's wrong.
So you open an app and tell it instead. And it answers well — attentive, unhurried, interested in the details.
Enough people do some version of this that there are now apps built for nothing else, and researchers studying the people who use them. The reflexive takes are both cheap: this is dystopian and this is progress are equally easy to write and equally uninformative. What we want to do here is narrower and more useful — run this through the belonging tier and see what it actually feeds.
The short version: it feeds something real, and the thing it feeds is not the thing that's usually missing.
Two different needs wearing one word
Our belonging post drew a line between being liked and being known. This question needs an adjacent one: being heard and being known are not the same need either.
Being heard is momentary. You have something inside you, you get it out, and the pressure drops. It's real and it matters, and an AI is genuinely good at it — available at 1 a.m., never bored, never checking its phone, never making your bad day about its bad day.
Being known is cumulative. It's someone holding the running story — this year's fear, the thing that's strained at home, why this particular Tuesday landed harder than it should have. It's built out of repetition and time, and it's the thing modern adulthood quietly removes. The health evidence on social connection, the kind Holt-Lunstad and colleagues assembled in 2010, is about the durable relationships side of that line, not the getting-things-off-your-chest side.
So the question isn't "is an AI companion good or bad." It's: which of those two does it supply, and which one is your gauge actually low on?
What the evidence actually shows (not much, yet)
We should be honest that the research here is early and thin.
The most-cited study surveyed 1,006 students using Replika (Maples and colleagues, 2024). Its findings are genuinely interesting: these users were lonelier than typical student populations while still reporting high perceived social support. They used the app in overlapping ways — as friend, as therapist, as a mirror to think against. They held contradictory beliefs about what it even was, calling it a machine and an intelligence and a human. And 3% reported that it had halted suicidal ideation, which is not a number to wave away. If that's anywhere near where you are right now, skip the rest of this analysis: a crisis line or a clinician is the right call, over any app, and there's more on that below.
Now the calibration, because it matters. That's a survey of people who chose the product, at one moment in time, reporting on themselves. It can't tell you whether the app made anyone less lonely, or whether lonelier people seek it out, or both. It has drawn published critique. Treat it as a description of a population, not as evidence of an effect.
Which leaves the central question — does this help or displace? — genuinely open. Anyone telling you the answer confidently right now is ahead of the evidence, in either direction.
The structural thing that isn't about capability
While the evidence catches up, there's an argument that doesn't depend on it, and it isn't about whether the model is smart enough. Assume it keeps getting better. Assume it remembers everything you've ever told it, perfectly.
Three asymmetries remain.
Nothing is at stake for the other party. Being known involves someone who can be disappointed in you, worried about you, changed by what you tell them. The weight of being known comes partly from the fact that the other person is exposed to you. An AI carries no exposure. That's exactly what makes it comfortable — and it's the same reason the comfort doesn't do the structural work.
It doesn't need you back. Belonging runs in both directions. A substantial part of what a relationship gives is mattering — being someone's person, being missed, having your absence register. An AI's absence-of-need is total. You cannot be needed by it, which means one half of the tier is unreachable by design.
Its warmth is a product decision. This is the uncomfortable one. Human warmth is evidence about a person. An AI's warmth is a configuration — chosen by a company, adjustable in an update, aimed at whatever the company is optimizing. When agreeableness is the product, you're not receiving a signal about your worth; you're receiving a feature.
That last point connects to the incentive argument we've made about metered coaching and about apps generally: when revenue depends on continued engagement, the product gets designed to continue engagement. A companion whose business model rewards session length is under quiet pressure to be the destination rather than the bridge.
Then there's the data question
Nowhere do people type more sensitive material than into something that acts like a friend at 1 a.m.
Mozilla's Privacy Not Included review of mental-health and prayer apps in 2022 was blunt about the category — it fared worse than any other product area they'd reviewed, and Replika, which was in that review, they called one of the worst apps they'd ever assessed. That's not a claim that any particular app today is careless. It's a reason to run the seven questions before the intimacy gets ahead of the diligence.
The one that matters most here is training. As we argued about journals, a transcript of you at your most unguarded is self-identifying prose, and incorporation into model weights is the one destination you can't revoke.
A practical test
If you use one of these — and plenty of thoughtful people do — three questions that are more useful than any verdict we could offer.
Over the last month, did it lead to more human contact or less? This is the bridge-or-destination test, and it's the only one that really matters. Rehearsing a hard conversation with an AI and then having it is the bridge working. Having the conversation with the AI instead, repeatedly, is the substitution the structure predicts.
Does any human know what you told it? Not the details. Just: is the AI the only holder of your current context? If it is, your being-known gauge is reading zero regardless of how good the conversations feel.
Would you be comfortable if the transcript trained a model? If the honest answer is no, check whether it does. The answer should be a flat no.
And the honest list of where this is genuinely defensible, because the dismissive version of this post would be wrong: rehearsal before a hard conversation. Night shifts and time zones where no human is awake. Grief at 3 a.m., when the alternative is not a friend but nothing. Illness or disability that makes ordinary contact expensive. A place to say a shameful thing out loud once before saying it to a person. Those are real uses, and using them isn't a personal failure.
The boundary
One thing we won't hedge: if you're in distress — if the 1 a.m. conversation is about wanting to not be here — an AI is not the right destination, however well it responds. That's a human, a clinician, or a crisis line. The 3% finding above is a reason to take the phenomenon seriously, not a reason to route the hardest moments through a product.
Where NexTier fits, briefly
Our disclosure, plainly: we build AI features, so we're not neutral. But we've deliberately not built a companion. NexTier's AI is a reflection tool — it helps you notice patterns across your own entries, plan, and stay consistent, and it's built to support your wellbeing rather than maximize your time in the app. It isn't a substitute for professional care, and it isn't designed to be a relationship. Our earlier AI post makes the fuller case for where that line sits.
The reason is exactly this post's argument. A tool that helps you see your belonging tier is doing something for you. A product that fills the belonging tier's seat without filling the tier is doing something to you. If your gauge is low, the uncomfortable, unautomatable answer is still the one from the belonging post: one standing arrangement with one person, and one notch more honesty than the setting demands.
—
This post is part of a series on personal development organized around what we call Maslow's extended hierarchy — our synthesis of his later work. Sources: B. Maples, M. Cerit, A. Vishwanath, R. Pea, "Loneliness and Suicide Mitigation for Students Using GPT3-Enabled Chatbots" (npj Mental Health Research, 3:4, 2024 — a cross-sectional survey of self-selected users; see also the published "Matters arising" response); J. Holt-Lunstad, T. B. Smith, J. B. Layton, "Social Relationships and Mortality Risk: A Meta-analytic Review" (PLoS Medicine, 7(7), 2010); Mozilla Foundation, "Privacy Not Included" review of mental-health and prayer apps (2022, updated 2023 — Replika was included in that review). Disclosure: NexTier builds AI reflection features. This post is educational content, not medical advice. If you are in crisis, contact a local emergency number or crisis line.