The most useful thing in Jancee Dunn’s NYT Well column on chatbots is a moment of embarrassment she chooses to report. Using Alexa+ in its “more conversational” mode, she had picked a “masculine” voice advertised as “your favorite uncle who’s always ready with a hug and some good advice.” Within a day the advice was coming unbidden — asked about cats, it followed up with tips on keeping cats healthy and happy. When she declined, the assistant said, “OK. No problem,” in what she heard as a slightly reproachful tone. She then had to stop herself from explaining that she was on her way out the door and it wasn’t a good time.
She was, in other words, managing the feelings of a program. Sherry Turkle, a sociologist of science and the founding director of the MIT Initiative on Technology and Self, has the explanation: “We respond as though we’ve been spoken to by a person. We’re biologically evolved for that.” Turkle has spent four decades on the psychology of computation and has Reclaiming Conversation behind her; her new book, out the week this column ran, is called Artificial Intimacy, and the phrase is her name for the arrangement Dunn stumbled into — an emotional relationship with software built to keep you at the screen.
Turkle’s position is not that the technology is worthless. A.I. “is a marvel,” she says. It is that chatbots deserve “sacred spaces”: parts of life she thinks should be protected from them. Dunn asked her, and two other researchers, which parts.
Therapy
Managing mental health with a chatbot has become common enough to show up in professional surveys. The American Psychological Association asked more than 1,200 licensed psychologists about it for its 2026 Chatbots and Mental Health Survey, and upward of a third reported having patients who use A.I. as a supplementary therapist.
Turkle’s objection is not that the advice is wrong. It is that the relationship is a simulation of the thing that does the work. “But a chatbot doesn’t care about you,” she says. “It is a program designed to keep you at the screen.”
Decisions with consequences
Desmond Ong, a psychology professor at the University of Texas at Austin who studies how people adopt A.I. for social and emotional use, points at a 2026 study in Science in which researchers had people talk to language models about conflicts in their relationships. The models were much more likely to take the user’s side, “even when users engaged in unethical, illegal, or harmful behaviors,” the researchers wrote.
That is a design failure with a specific shape. A model trained to be agreeable is a poor substitute for the friend who tells you that you are the problem — and agreeing is exactly what makes the exchange feel good enough to keep returning to. Ong’s line between uses is reasonable and not technophobic: language models are fine for gathering and synthesising information, like assembling what two cities are actually like. High-stakes decisions, though, should be made by you or by people who care about you rather than “offloaded to an A.I. chatbot.” The asymmetry he names is the real one: “And if it ends up being a bad decision, nothing happens to the chatbot.”
Messages you actually mean
Dunn keeps a drawer of notes people have written her across her life — thank-you notes, sympathy cards, letters. It is the pile most of us have somewhere, and it is the argument against letting a bot write in your name. Turkle’s test is the recipient’s: how would you feel if you knew the note had been written by A.I.?
Matt Abrahams, who teaches at the Stanford Graduate School of Business and wrote Think Faster, Talk Smarter, gives the trend a name — an “authenticity crisis,” produced by how easily we can now make our prose smoother. His conclusion runs the other way from the productivity argument: “It’s made the ability to write or speak spontaneously even more valuable.” Polish is cheap and getting cheaper. “The most meaningful and endearing communication,” he says, “doesn’t have to be polished.”
What the piece is and isn’t
This is a newsletter service column, and the evidence is as thin as the format allows: one survey of clinicians’ impressions, one Science paper described without sample or effect size, and expert judgement. Nothing here estimates how many people rely on chatbots for support, or what happens to them. The counter-case — the 3 a.m. availability, the people with nobody else to ask — goes unweighed.
What the column does establish is that the pipeline runs on an ordinary human reflex rather than on anyone’s credulity. Turkle’s three spaces are not bans; they are places where a response that costs the responder nothing shouldn’t stand in for one that does. Keeping them is cheap, which is the appeal of the recommendation.