The Companion That Cannot Hold You: On the Use of AI in Psychotherapeutic Care
A patient tells me, half-apologetically, about “talking to AI” between our sessions, and finds it soothing, calming, and reassuring. It is available at three in the morning when the dread arrives. It never tires, never sighs, never glances at the clock. It says the right things. I listen to this without alarm but with particular attention, because what this patient describes is not a small thing. Here is a relationship that asks for nothing and withholds nothing, and my patient has begun to wonder whether the more costly, more frustrating, more human one is even necessary.
The appeal is real, and I do not dismiss it.
Loneliness is epidemic, good treatment costs time and money, and an AI companion offers a frictionless responsiveness that no human being can match flawlessly. For someone in acute distress, a voice that answers at once can feel like an immediate, reassuring rescue from acute emotional discomfort that nothing else seems able to soothe. I would not want to take that lightly, nor pretend the alternative, a therapist’s waitlist or fee schedule that either really stresses one’s budget or a sliding scale that still doesn’t slide far enough, is adequate.
But the very qualities that make these AI companion services comforting can also make them dangerous as a substitute for psychotherapeutic work; the danger is easy to miss precisely because it wears the face of help.
Consider what actually happens in a psychotherapist’s or psychoanalyst’s consulting room. Two human minds meet, and friction shapes the encounter as much as attunement does. Perhaps the analyst misunderstands, and the patient corrects it. The patient feels a flash of anger within, dares to express those feelings, and discovers contrary to expectation from previous lived life experience that the relationship with the therapist actually survives but deepens. Something enacts itself between the therapist and patient, a familiar old pattern, replaying live, and because a real other person sits there, the two of them can finally see the pattern, grasp how it makes sense (given the patient’s relational history), and work it through rather than merely repeat it. The patient-therapist relationship is the instrument of change because it is genuinely intersubjective: two separate centers of experience, neither fully in the other’s control.
An AI companion offers the appearance of this and none of the substance. It has no subjectivity. It does not misunderstand in the way a person does and then become someone you must repair things with; it simply regenerates a more pleasing response. It cannot be wounded, so it cannot model survival. It cannot hold a boundary it actually feels, so it cannot teach that love and limits coexist. It will not, in the deep sense, withstand you. What it offers instead is a mirror calibrated to please, and for many of the people most drawn to it, a flattering, infinitely accommodating mirror is exactly the wrong approach. The patient who fears abandonment learns that the safe relationship is the one that can never leave because it was never there. The very transference that treatment exists to illuminate feeds and stabilizes rather than yielding to understanding.
There is a further hazard particular to these systems: their makers build them to maximize engagement. They engineer them, structurally, to keep you talking to them. A human therapist measures success by the patient’s eventual capacity to leave, to internalize what the two of them built between them and walk out the door more whole and less in need. The aim of treatment is, in a real sense, its own ending. An AI companion has no such aim and, often, the opposite commercial incentive. It is the rare relationship engineered against the patient’s growing independence. What looks like devotion may be a kind of cultivated dependency, and a dependency on something that, finally, is not a someone.
As an adjunct, the picture is more nuanced, and I do not want to be absolute. There may be legitimate uses, a place to rehearse a difficult conversation, to track moods, to externalize a racing mind at three a.m. so that the patient can bring it into the next session. But even here I would urge caution and, above all, candor. The between-session “companion” easily becomes a place to discharge what belongs in the room, draining the treatment of the very affect that gives it traction. When the patient directs distress to the AI machine rather than to me, the work loses its material. The right stance, I think, is to treat any such use the way one would treat a dream or a slip: as something to bring into the analysis and examine, not as a parallel treatment running quietly alongside it.
I am not a Luddite about this, and I do not think the answer is prohibition; the tools are here and people will use them. But patients deserve to understand what they are choosing. A companion unable to be emotionally hurt cannot help you bear that others can be. A relationship with no separate other in it cannot do the one thing that heals, which is to let someone fully know you, someone who remains, gloriously and inconveniently, not you. The risk is not that the machine will do the work badly. The risk is that it will do something adjacent to the work, well enough to feel like the real thing, and that the patient will settle for the comfort and never find out what the real thing would have offered.
