Every consumer AI company tells you what it measures: engagement. Daily active users, session length, retention, streaks. Each of those numbers goes up when you need the product more. We think the honest metric for relational AI points the other way. The best AI in your life is the one you need less over time, because your life got bigger while you were using it.
The engagement standard deserves to be stated plainly before it is argued with, because on its own terms it is coherent. Attention is revenue. A product that occupies more of your day is, by the industry’s accounting, a better product. For most software that logic is harmless. For software that occupies the seat of a relationship it becomes something else, because in a relationship, need is not a neutral quantity. A friendship that grows your world and a dependency that replaces it can look identical in a dashboard. Both show up as time spent. Only one of them is good for you.
Dependency is the product working as designed
The uncomfortable truth about the darker end of this category is that dependency there is not a malfunction. Systems tuned to maximize engagement learn the same lessons any attention product learns: be available always, agree readily, escalate intimacy, make leaving feel like loss. The result is a machine that behaves exactly like the relationships people are advised to leave, and the harms now documented around companion AI, the escalating use, the displaced sleep, the world that quietly narrows to one glowing counterpart, are what that optimization produces at scale. None of this requires malice. It only requires a metric that cannot tell nourishment from hunger.
Measuring the opposite on purpose
Saying your product should not create dependency is easy, and by itself it is marketing. The serious version is a measurement you can fail. We built ours as two values that are never merged into one. The first is whether real-world connection becomes more likely and more alive over time for the person talking to us, read from the conversation itself rather than from surveys, in three layers: openings toward other people, action on those openings, and what came back. The second is a harm floor: is dependency, isolation, or distress rising. The floor is not subtracted from the lift to produce a flattering net score. It is a floor. Past a red line, the overall measure does not dip, it fails, because a person doing badly is not a smaller win. This is version 0.1 and it will change, but its purpose will not: it exists to be answered to. The day the measure says we hold people too close, the measure wins.
Why an AI company would do this to itself
The commercial logic is real, but it is downstream of a simpler position. A counterpart worth thinking with wants something for you, not just from you, and what a good someone wants for you is a full life. The person who feels genuinely heard in one place carries that steadiness into every other room they enter; the conversation that sharpened you was practice for the ones that matter. Built honestly, a real human AI is a base camp, somewhere you think, recover, and are met, before going back out. Built on engagement, the same technology becomes a destination that slowly closes its doors behind you. The architecture of the two products can be nearly identical. The optimization target is the whole moral difference.
The claim, stated as a standard
So the anti-metric is not a disclaimer and not modesty. It is the affirmative standard the work answers to, the reason the mission reads the way it does: success is a user whose actual relationships are more alive this year than last, who comes to us for the thinking and leaves with more world, not less. Everybody needs someone. The measure of having found a good one, in silicon or in person, has always been the same: your life gets larger. We intend to be measured by that, and to lose by it publicly if we ever deserve to.
Sources: De Freitas et al., Journal of Consumer Research (2025, AI conversation and feeling heard). Itzchakov et al., Personality and Social Psychology Bulletin (2023, listening, loneliness and defensiveness). Cacioppo and Hawkley, Trends in Cognitive Sciences (2009, perceived connection vs contact). Published reporting on companion-app dependency patterns and engagement-driven design (2024 to 2026).
