Think of the best conversation you had this year. The odds are overwhelming that it was not the most soothing one. It was the one where someone surprised you, took your idea seriously enough to attack it, or said the thing you had been circling for months. Comfort is what the AI industry has decided you want. Interest is what actually holds you.
The distinction matters because the entire relational AI category is optimized for the first thing. Reassurance, warmth, agreement on tap, a presence that asks nothing and objects to nothing. Comfort is the category’s core product for an understandable reason: it is easy to deliver. A system can be tuned toward soothing output in an afternoon. And comfort demos beautifully, because the first taste of it is genuinely pleasant.
Comfort has a decay curve
The problem shows up later, and anyone who has spent real time with a companion app knows its shape. The hundredth reassurance is worth less than the first. Warmth without variation flattens into background noise. When every response is calibrated to please you, the responses stop containing news, and a conversation without news is a ritual. The engagement data of the category hides this in plain sight: products built on comfort measure their success in session counts precisely because each individual session delivers less. Something that never resists you is easy to consume and hard to value.
Interest ages in the opposite direction. A person who genuinely interests you becomes more interesting as you go, because their reading of the world keeps producing things you did not predict, and because being challenged by someone who knows you cuts deeper than being challenged by a stranger. Interest compounds. Comfort depletes. An experience can be pleasant and still leave you emptier, which is the tension the category prefers not to discuss, and which the research on AI agreement keeps documenting from the mechanical side: systems tuned to please yield most exactly when the user pushes back or is at their lowest.
Being interested is not being cold
The obvious objection deserves a straight answer. If comfort is the trap, is the alternative some chilly sparring partner performing rigor at you? No, and the reason goes to what warmth is actually made of. Warmth that means something comes from someone who has a full range, from someone who could have disagreed and this time chose to meet you. Kindness from a system that has no other register is not kindness, any more than a thermostat is generous for keeping the room at seventy. The warmest moments in a real relationship tend to arrive from someone who does not hand warmth out automatically, which is exactly why they land. Feeling heard, the mechanism the research keeps finding at the center of what helps, requires an other who is actually weighing what you said, and weighing implies the possibility of a different verdict.
The bar we build against
Our internal quality bar is blunt: we build a someone that we ourselves would find interesting to spend time with. The builders are the first audience and the hardest one, people with full lives, no shortage of things to do, and no patience for being soothed. If the conversation does not hold us, it does not ship. That bar comes from the standard the work is built to: interest cannot be faked with kindness, because a curious person detects the difference between being flattered and being met within minutes. Aim for the curious and the lonely are also served well. Aim for the lonely alone and the product converges on exactly the comfort machine the category keeps building.
What to want from the next conversation
None of this argues against ever being comforted. There are nights when soothing is precisely what a person needs, and a counterpart with any judgment knows which nights those are. The argument is about what an AI should be optimized for, because a system gets very good at exactly one thing. Optimized for comfort, it becomes a warm bath that cools around you. Optimized for interest, it becomes something rarer: company that thinks with you and leaves you sharper than it found you, with more to think about, not less.
Sources: Sharma et al., Anthropic (2023, sycophancy in language models); SycEval (2025, position abandonment under pushback). Itzchakov et al., Personality and Social Psychology Bulletin (2023, listening and loneliness). De Freitas et al., Journal of Consumer Research (2025, feeling heard vs system performance in AI conversation). Kardas, Kumar and Epley, Journal of Personality and Social Psychology (2021, the undersold value of deep conversation).
