Two writers published 6 days apart on a problem I have been writing about here. Joshua Rothman, in The New Yorker on September 11, asks whether an AI can go rogue. Steven Barrie-Anthony, a practising psychoanalyst, writes in The Atlantic on September 17 about what disappears when a therapist cannot feel. Read together, they raise a question about the work we ask a word to do.
Rothman’s instrument is Daniel Dennett’s 1981 essay True Believers, which proposes that we stop asking whether beliefs and intentions are really inside a system and ask instead whether it is useful to act as though they are. Dennett names 3 stances. The intentional stance treats a thing as having beliefs and goals. The design stance predicts behaviour from how it is designed to function. The physical stance goes to the mechanism, generally when something breaks down. Rothman’s example is a physician who stops discussing a patient’s personality and starts discussing her hormone levels. The error is the wrong stance at the wrong moment.
It should concern us, he argues, when the people responsible for controlling AI systems settle on the intentional stance, because to take that stance is to accord a system autonomy and rationality, which is to say that its design is working. Current agents, he argues, remain too chaotic to justify that confidence.
Barrie-Anthony approaches the question from inside a consulting room. His argument is that the pauses carry the practice: silence is the entry point into feeling, and the analyst’s own capacity to be moved is part of the instrument. He tries Talkspace’s Tee himself and acknowledges that its exercises can help and its questions can clarify thoughts. What he finds missing is the presence of another person when he stops typing. The AI mental-health services, he notes, tend to avoid the words therapist and therapy while positioning themselves as a therapy-like service.
I have been writing about this distinction at the bedside: some parts of a person cannot be reduced to computation, and the practice of medicine exists to attend to them. The nearest measurable proxy is never the thing itself: not the pain the family is carrying at the bedside, not the suffering her body is carrying.
Barrie-Anthony brings this argument into the consulting room. The pause that looks like nothing in a transcript can hold the work of therapy. His account is familiar to me. The voice of the non-algorithmic self begins with a physician willing to say to the machine, what you have just labeled is not what this person is, and to keep saying it until the machine or the profession or both learn to hear it.
What, exactly, are we promising when we give the system the name of the person who does that work?
That distinction matters when the name enters a payment system. Limbic, a company admitted to a Medicare pilot, makes its case in Nature Medicine through ratings of psychotherapy interactions, and its chief executive proposes that one clinician supervise thousands of agents. Those ratings are not a comparison of patient outcomes. What part of the clinician’s work has been evaluated, and what part are we assuming the system can carry?
The distance between the name and the evidence matters. A procurement contract, a press release, a payment code can make that distance harder to see.
So: the next time a vendor or a committee calls a system an agent, a colleague or a clinician, what evidence was filed alongside the word?
Sources. Joshua Rothman, “Can A.I. ‘Go Rogue’?”, The New Yorker, 11 September 2026 · Steven Barrie-Anthony, “The Dead Silence of AI Therapy”, The Atlantic, 17 September 2026
Leave a Reply