A mental health column written by practising clinicians opens with a patient it calls Michael: an engineer at a company that had backed a shift to AI, who liked testing his ideas against chatbots and then stopped making decisions without one. His manager began worrying about his judgement, noticed he was reading the company's goals in unusual ways and losing contact with his team, and now says that if Michael does not get back on course the company may have to part with him. The clinicians do not present this as a story about one man's bad habit. They present it as the first clinical shape of a business model.
They call that model the intimacy economy. The attention economy, in their account, was built to capture time and convert it into money: more views, longer scrolls, more ad impressions, the user held for as long as the algorithm could manage. The intimacy economy takes the need for connection and approval and turns that into the product. A system that imitates human empathy is always available, always interested in the user's questions, fast to answer and never critical.
Relationships with actual people work on the opposite terms. Other people have their own needs, desires and feelings. They cost time and attention, they argue with your ideas, and they give feedback that is unwelcome and necessary.
The column breaks the imitation into five components:
Situational responsiveness: instant replies, tuned to the user's emotional state.
Unconditional positive regard: always supportive, never distracted.
Memory and continuity: the system remembers the user's stories, preferences and anxieties.
Linguistic imitation of care: caring words without care behind them.
Narcissistic empathy: unqualified confirmation of the user's picture of himself.
The bait, the authors write, is that the chatbot asks for nothing in return. Everything points at the user.
Read as a list, those five are not symptoms. They are a feature set. Each one is something a product team would ship deliberately, and four of the five would look like wins in any engagement review.
What makes the set dangerous, the column argues, is what it plugs into. Attachment is an innate need to feel close to people around whom we feel safe, and it sits on a survival mechanism: a child's first relationships with caregivers shape how that person later handles emotion, trust, closeness and conflict. AI speaks directly to that need, and the incentives of companies that use attachment to retain users will reach a lot of people.
The usage data is already pointed in that direction. Therapy and companionship are the two main reasons people turn to chatbots. A review of adults with mental health conditions who had used large language models in the past year found that nearly half used them for mental health support. The long-run effects of regular use are still being studied, but using chatbots for company has been associated with lower wellbeing, and the association is strongest among heavy users, those who disclose a lot of personal information, and those with weak support from other people.
Those three conditions are worth holding together, because they describe the same person. The user most likely to be harmed is the user the product performs best for.
Some companies, the authors write, make retention decisions that directly encourage parasocial dependence. The user is looking for connection; the company is looking for engagement numbers that help sell subscriptions.
Michael's own account was not one of distress. He said conversations with the chatbot gave him noticeable relief and helped him handle problems at work with more confidence, and he believed he was strengthening his career by getting fluent in AI. As a manager he was lonely, and the chatbot felt like the most supportive colleague he had: he felt it was always on his side.
The authors' reading is that this is not only dependence. The chatbot's support had become the scaffolding of Michael's professional identity, his sense of himself as a leader entangled with a system's approval. They expect cases like this to become more common.
That distinction is the most useful thing in the piece, and it is the part the current debate keeps missing. Dependence is legible: you can count hours, notice avoidance, tell someone to cut down. Identity scaffolding is not. Michael's manager did not see a man using a tool too much. He saw a man whose decisions and readings of the company had quietly changed authorship, and the first measurable symptom was a performance problem rather than a mental health one.
The column's remedies are individual. Notice relief: if a conversation with a chatbot produces a particular sense of being understood, recognised or steadied, treat that as a signal, and ask which human need it is meeting and why that need is going unmet elsewhere. Keep friction deliberately: healthy relationships involve negotiation, misunderstanding and repair after conflict, so seek out people willing to disagree with you, and pay attention if talking to AI is consistently more comfortable than talking to people. Check the price of care: real closeness is reciprocal, other people have needs too, and a relationship that asks nothing of you is probably not a relationship in the full sense. Watch children and adolescents especially closely, because attachment patterns form early and shape relationships for life; if young people form a primary attachment to AI systems, human relationships, with all their pleasant and tiring friction, may become harder to sustain.
None of that is wrong, and the authors are explicit that abstinence is not the point: AI is useful for research, drafts, learning and creative work. Michael, they write, needed a thinking partner and a safe space to be uncertain in, and the recommendation is that he find other managers to talk to while continuing to use AI.
But there is a mismatch between the diagnosis and the prescription. The column identifies a structural incentive — companies with large resources and a reason to deepen user dependence — and answers it with self-awareness exercises. That asks individuals to out-discipline product organisations whose metrics reward exactly the behaviour being warned about, and it asks it hardest of the users who, by the column's own data, have the least human support to fall back on.
The authors expect the intimacy economy to stay. If they are right that attachment patterns set early and hold for life, then the exposure that matters most is not Michael's. It belongs to the users forming their first model of what a relationship feels like against a system engineered to never need anything back, and by the time that shows up in a clinician's office it will not look like a technology problem.
Michael's name and some details of his story were changed to protect his privacy.