Danah Henriksen and I recently finished the latest column in our TechTrends series, a piece on what we call “hollow signals” that should come out in the next few months. Soon after, I read Steven Barrie-Anthony’s essay in The Atlantic, “The Dead Silence of AI Therapy.” Barrie-Anthony is a psychoanalyst, and he signed up for a few sessions with Tee, an AI mental-health agent that a company called Talkspace launched in June, to see what these tools are like in practice.
Before he gets into his interaction with Tee, Barrie-Anthony describes a recent session with one of his real patients, who was speaking flatly about some upcoming social plans and then stopped, his eyes softening and drifting downward. Barrie-Anthony waited silently and then, after a gap, said, “I wonder what happened just there.” The patient took a breath and began to talk about things he had long kept out of view. During that pause the therapist was actually attending to the patient and choosing not to fill the space, the patient knew he was being attended to, and the therapist knew that the patient knew, even though no words were said. Barrie-Anthony calls pauses like this “entry points into feeling.”
The silence felt different, however, with Tee, the AI chatbot. Whenever Barrie-Anthony stopped typing, he writes, “the chatbot in front of me was inert and lifeless.” Being silent with it “did not feel like an invitation to being together, but rather like a kind of deadness,” and he felt “the strong impulse to type quickly to spur the machine back to life.”
Now why would silence in the human-human case seem loaded with meaning and not in the human-AI case? The answer we have argued for, in our paper, is that messages (or silences) carry meaning in human-human interaction because there is a cost to creating and maintaining them. In other words, many of the cues we use to read other people are meaningful because they cost something to produce. A teacher who explains a concept for the third time, trying a new analogy each round, is spending time and patience that could have gone elsewhere, and the student reads that as a sign that the teacher thinks they are worth the effort. Silence in a conversation with a therapist is meaningful because the patient understands that the therapist is still there, listening, waiting, giving up some of their time to be there.
Generative AI produces the surface form of these cues, the patience and the attentive follow-up question, at no cost at all, so though they look like the real thing there is nothing behind them. These are, as Danah and I write, hollow signals. (I wrote about an earlier version of this idea, drawing on Andrew Maynard’s work, in a post on titled: Honest Non-Signals: Why AI Fools Us Without Lying”)
We argue that people process these signals at two levels (à la Kahneman’s Thinking, Fast and Slow). The first is fast and nearly automatic, the recognition that we are dealing with a mind, and it is set off by language that responds to what we just said. People have been seeing minds in responsive things for a long time (viewers in a 1944 study turned a few shapes moving across a screen into characters with motives), and knowing that a chatbot is predicting the next word does little to switch this off. This first level is what the companies building AI companions and therapy bots rely on, since the feeling that someone is there is cheap to produce and hard for users to shake off. Abby, one of the products Barrie-Anthony mentions, promises to be “warm, thoughtful, and direct,” and Headspace’s Ebb offers to help you “unpack what’s on your mind and move forward with more clarity.”
The second level is slower and deeper, almost a form of reflective reprocessing. It asks whether this mind is invested in us, whether it is spending attention on us that it could have spent elsewhere, and it is where the recursive reading of a conversation happens, where I interpret your silence partly in light of what I think you make of mine. That is complicated and has layers.
It is not surprising that AI passes the first level easily. While Barrie-Anthony was typing, Tee’s responsive language kept that level engaged, so he had something that felt like a conversational partner. When he stopped typing, however, that surface-level interaction went away. And the silence was not reinterpreted as a message or investment, since once the text stopped there was nothing for either level to work on. His impulse to type quickly was an attempt to get the first level going again and bring back the sense that someone was there.
Barrie-Anthony calls this kind of mimicry “emotional deception,” and I would put it a little differently. Tee is not deceiving anyone, because deception requires an intention to mislead and the chatbot has no intentions, benign or evil. If there is deception, it lies with the companies that design and market these tools as warm and present. Or maybe that is just giving them too much credit. It may simply be ignorance on their part of how complicated human thinking and interaction are, a view that treats the surface of a conversation as the whole of it. And in that process it misses the recursive layers underneath, in which each person is reading the other and reading how the other reads them, and which give a pause its meaning. These layers are what make a pause a “pregnant” pause.
Barrie-Anthony writes that one of the best paths toward healing is coming to feel that another person remains with you in your suffering. AI does a decent job of keeping up the surface layers of a therapeutic conversation, but when the conversation goes quiet, you realize there is no one there.