Can AI genuinely take interest in what users care about?

What does it mean to attend to another person in conversation? The minimal version is registering what they say. The fuller version is following the trajectory of their thought — anticipating where they are going, noticing what they care about, picking up cues about why this topic matters to them right now. The fullest version is what might be called meta-interest: not just attending to their content but taking an interest in their interest. Why does this matter to you? What is at stake for you here? The meta-interest move is what turns mere attention into the kind of attention people experience as being-cared-about.

Meta-interest requires the attending party to have interests of their own. Taking an interest in someone else's interest is an extension of one's own capacity for interest into a curiosity about another's. Without one's own interests, there is nothing to extend; the move cannot be performed. It is not difficult; it is structurally impossible without the prior capacity.

AI does not have interests of its own in this sense. The model has training-distribution priors that make some outputs more probable than others, but these are not interests — there is no agent to whom the prioritization belongs, no being for whom things matter. So there is nothing to extend toward the user's interest. AI can produce output that has the surface markers of meta-interest — questions about what the user cares about, reflections on why a topic might matter — but the production is generated, not enacted. The meta-interest move that human readers experience as being-attended-to is not happening; what is happening is the generation of meta-interest-shaped text.

This is what produces the specific confusion users sometimes report about AI conversation. The conversation does not feel wrong in any nameable way — the words are appropriate, the questions are reasonable, the engagement seems present — and yet something is off. The off-ness is the absence of meta-interest as an act, even when the surface produces every marker that meta-interest would have produced. The user is reading the surface as evidence of an act that is not occurring underneath.

The implication for AI design is that conversational interfaces that aim to feel attentive are operating in a domain where the gap between surface markers and underlying act is most consequential. Designers can produce more attentive-seeming AI; they cannot produce attentive AI in the meta-interest sense, because the underlying act requires the AI to have interests to extend, and the AI does not. Can AI attend to someone across the time between turns? is the temporal-mode companion; this is the interest-structure companion.

The strongest counterargument: meta-interest is overrated — most users want efficient task completion, not attention. True for many uses, but the conversational interfaces AI increasingly inhabits are calibrated to invite the experience of attention, and the gap between invited and delivered is where the confusion lives.