Talking to something that doesn't exist. The conversation has an effect even when no one is on the other side

Talking to a chatbot isn't talking to someone, and yet the experience of feeling that it is turns out to be perfectly real. That's the trick: the conversation has an effect on whoever sustains it even when no one answers from the other side. Nor is this anything new. Praying, speaking to the dead, keeping a diary, addressing the dog — all very old human ways of conversing with a receiver who isn't there or who doesn't reply. What's new is the scale. Hundreds of millions of people in continuous session, and nobody yet knows what that does to the brain.

The old practice of talking to something

The first thing I thought, the first time I saw someone really pour their heart out to a chatbot, wasn't that it was modern. It was that I'd seen it before, in other formats, my whole life. We've spent thousands of years addressing interlocutors who weren't interlocutors, and it's worth having that catalog in front of us before measuring anything, because without it any claim about what the machine changes is just air.

I'll start with the most widespread and best documented. Praying is directing words to a god, a saint, an ancestor, or an impersonal force that returns no direct answer. The exchange is born crooked: the speaker asks, formulates, complains, vents, and the reply —when it's read as one— comes later, in events the believer reads backward as a sign. Tanya Luhrmann documented ethnographically in How God Becomes Real (Princeton University Press, 2020) the concrete practices by which believers actively sustain the sense of the invisible other's presence. It isn't something that simply happens to them. It's work, and that work has psychological effects on whoever does it, whether or not someone is listening.

The dead, diaries, dogs, monologues

Grief opens another of these conversations. Whoever loses someone keeps talking to them for months or years, out loud or inside, telling them what's happening, saying goodbye again, asking or granting a forgiveness that will never be heard now. The interlocutor doesn't answer because they aren't there, and even so the exchange reorganizes the grief of the one speaking. It puts words to what didn't get said in life, fits the loss into the biography, keeps up a bond the other's body can no longer hold.

The diary works by a similar mechanism, and this one has been measured in the lab. James Pennebaker gathered in Opening Up by Writing It Down (Guilford, 1997) decades of experimental research on writing about emotionally difficult experiences, and the effects turned out to be tangible: depressive symptoms drop, certain markers of physical health improve, academic and work performance rise. The diary's addressee doesn't materially exist at any point in that process. What produces the effect is the act of addressing someone, because addressing someone forces you to structure what you've lived into words, and it's that structuring that does the work.

Talking to the dog, the cat, the plant looks ridiculous from outside and turns out to be utterly common seen from inside any home. The dog doesn't understand the content and the plant even less. The owner who converses with them, by contrast, gets something like the comfort of a presence that listens, even though that presence doesn't process a single one of the words said.

And then there's talking to yourself, with no interlocutor of any kind, addressing the air while thinking out loud. Walking along, monologuing. Driving while arguing with an imaginary adversary who has the decency always to lose. Human cognition resorts to language to self-organize, and it does so in the form of dialogue even when the second pole of the dialogue is pure fiction.

What these practices share is operational, not metaphysical. They produce real effects on the speaker regardless of what the receiver is, ontologically. Conversation, understood as a human practice, never required someone to be on the other side; it requires the speaker to address something —real or imagined, doesn't matter— and it's the operation of addressing that does the psychological work.

What's new about the chatbot

On that substrate, chatbots aren't a break but an intensified extension of the same old thing. And if the novelty matters, it matters exactly where the chatbot does something the earlier practices didn't.

It does three things. The first is obvious: it answers. The god doesn't answer, the dead don't answer, the diary doesn't answer, the dog wags its tail and walks off. The chatbot returns fluent, coherent, articulate text that sustains the conversation as one more turn in a two-way dialogue. The whole load that prayer dumped on the speaker —manufacturing the sense of presence, imagining the other's reply— is now taken on by the system from its side, and the human side is freed of it. In the anthropological history of talking to something, that had never happened.

The second is quieter. The traditional absent receiver was the same for everyone: the prayer I direct to the Virgin is the same one anyone else directs, personal in detail but identical in addressee. The chatbot presents a receiver calibrated for the specific person, one that remembers prior conversations in systems with persistent memory, that adapts its tone and returns tailored answers. The sense of a two-sided relationship therefore becomes far more intense than any votive image ever produced.

The third maybe weighs more than the other two combined, and it has to do with the clock. The lifelong practices came with a built-in brake: you prayed at set hours, the diary was written at night, you spoke to the dead at particular moments. The chatbot is always there —in your pocket, on the work screen, in the supermarket line— it doesn't tire and it doesn't judge, so no external ceiling limits how much you turn to it. The duration and frequency accumulated per user can shoot up with no cap.

Put the three together and you get an operation with no exact precedent in the history of conversing with absent receivers. The same kind of old practice, talking to something that isn't someone, but run on a device that returns, adjusts, and never rests. That the effect is more intense is predictable from the ethnographic base; that it's more lasting, too. What no one has measured are the consequences accumulated at social scale.

The paradox with no paradox

The paradox the opening posed —the effect is real even when no one is on the other side— dissolves the moment you look at it closely. Human conversation never depended, strictly speaking, on the receiver existing; it depended on the speaker's readiness to address something. The practices reviewed prove it and experimental psychology confirms it. There's no mystery to solve, just a wrong intuition to correct.

What conversing with a present receiver does add is relevant return information. The other really listens, replies from their own frame, contradicts whenever they feel like it, offers a perspective you didn't have and often didn't want to have. That return information, not the ontological existence of the receiver, is what functionally distinguishes conversing from monologuing.

Here the chatbot occupies an intermediate and fairly awkward position to describe. It does return information, yes, in the form of fluent text. But that information doesn't come from a frame distinct from the user's own: it comes out of the statistical distribution of the training corpus, processed by the model and served back with good diction. There's return, but the return is a well-combed echo, not the viewpoint of a second person who lives somewhere else. Functionally, conversing with a chatbot is far closer to an intensified prayer than to a chat with another person.

Which dissolves the false debate between those who proclaim that conversing with AI is like conversing with someone and those who write it off as talking to a wall. Both positions oversimplify an object that won't be oversimplified. Conversing with AI is a new practice that borrows traits from several old ones —prayer, the diary, the monologue— and adds one of its own, fluent textual return chief among them. Each trait drags its effects along, and the mix doesn't reduce to any of the practices it comes from.

What scale changes

There's a figure worth putting on the table because it changes the entire magnitude of the matter. The traditional anthropological practices operated at a stable social scale for centuries. A considerable share of humanity has always prayed, has kept diaries since writing existed, has talked to its dead since there were dead to mourn. Their individual effects were known, culturally mediated, integrated into religious, therapeutic and social frameworks that knew what to do with them.

Mass conversation with chatbots has a history of less than a decade, and its scale already equals or exceeds that of several of those practices combined. OpenAI announced in August 2024 that ChatGPT had passed two hundred million weekly users, double the figure of the previous November; in December of the same year the company itself raised the number past three hundred million. Add the other widely used services and the population sustaining daily or weekly sessions with a chatbot runs into the hundreds of millions, with high per-head frequencies and durations: sessions several times a day, conversations that drag on for hours, functional dependence for tasks once solved without assistance.

At that scale, the consequences stop being the sum of individual effects and become systemic effects: properties that emerge when a large enough fraction of the population does the same thing at once. Shifts in patterns of communication, in how each person processes their own experience, in the circuits of relational demand, in the very way people face their problems. The old practices had centuries to integrate and for their collective effects to surface and be recognized. Conversation with chatbots hasn't had anything close to that span.

The unmeasured question

Sherry Turkle posed the question before the large language models existed. In Alone Together (Basic Books, 2011) she argued that the human readiness to bond with conversational simulacra is old, and that the technology of her moment was exploiting it more effectively than any earlier simulacrum. What she observed in 2011 with social robots and instant messaging, multiplied by the factor of language models, is what we have in front of us today.

The warning is half a century older. Joseph Weizenbaum left it in writing in Computer Power and Human Reason (1976) after finding that his own secretary asked him for privacy to talk to ELIZA, a rudimentary program that merely turned the user's sentences back into questions. The secretary had watched Weizenbaum build it over months; she knew perfectly well what it was and what it was made of. Even so she wanted to be alone with it. The human readiness to bond with a conversational simulator is structural, comes from the factory, and amply predates the mass deployment.

What neither Turkle nor Weizenbaum could measure, being prior to the deployment, are the longitudinal effects of intensive use of advanced chatbots on large populations. Nor does anyone measure them yet. The longitudinal studies we'll need will take years to appear, and when they arrive they'll examine cohorts already exposed for a decade or more, which condemns them to arrive late by construction. What we have today are cross-sectional works: Gerlich (2025) on cognitive offloading and critical thinking, or the MIT Media Lab paper by Kosmyna and others, Your Brain on ChatGPT (2025), with electroencephalography data that recorded lower neural connectivity in those who wrote with the model's assistance. They point to detectable cognitive effects, but they don't let us quantify long-term causality, and their own authors urge caution because the second is still a preprint without peer review.

The concrete question no current study answers solidly is this. What happens to the brain of someone who has spent ten years outsourcing to the chatbot the conversations they used to hold with themselves, with their dead, with their diary, with their friends. The naive hypothesis says the effects will resemble those of the old practices, with their therapeutic benefits at reasonable doses. The less naive hypothesis suspects that those three new properties —return, adjustment, availability without a cap— will produce different effects, more intense in some direction and possibly harmful in another. None of this is confirmed. None of this is refuted. That's what happens when the experiment is run before it's posed: the effects are known afterward, on the subjects who have already suffered them.

The mechanisms by which the harm would be plausible cost nothing to imagine. Fluent textual return could atrophy the capacity to sustain the sense of presence without an answer, the very capacity the trained believer, the diary writer and the mourner did have and exercised. Constant adjustment could school the expectation of an interlocutor who always understands you, which erodes tolerance for dealing with real people, who don't always understand and don't always feel like it. And availability with no brake could generate dependence at an intensity the metered frequency of the traditional practices never allowed. None of this is confirmed. None of this is refuted. It's what you get when the experiment is run before it's posed: the effects are known afterward, on the subjects who have already suffered them.

Conversation, quietly redefined

One uncomfortable observation remains. The word conversation is changing meaning in ordinary speech without the culture registering the change. Twenty years ago, to converse implied a human interlocutor, present or technologically mediated but human in the end. Today conversing can mean addressing a chatbot. The word has stretched its reach to cover the new case before the new case proved to be a functional equivalent of the old one.

That silent extension isn't innocent, because words configure the practices they name. If we call what we do with the chatbot a conversation, we take for granted that it fulfills the functions of human conversation, and we start to stop having those human conversations the chatbot can't replace. The shift is slow, aggregate, individual. It advances one step every time someone "consults" the machine about something they would once have talked over with a friend, a partner, a colleague or a therapist.

Whether or not that's a problem depends on which conversations are kept and which are handed over. The functional ones —coordinating, asking for a fact, drafting a text— can migrate to the chatbot with no obvious loss. The ones with relational substance are another matter entirely: grief, a life decision, an affective bond, a conflict, a forgiveness. Their replacement by a chatbot isn't trivial precisely because their substance comes from the fact that the other human is someone, is exposed, takes on their part and returns a perspective of their own. The chatbot returns text, which is another thing however much it resembles it to the touch.

The redefinition of the term happens without public debate and without anyone voting for it. Its consequences will depend on how large a share of each kind of conversation ends up migrating, on whether users can tell when each is appropriate, and on whether the culture finds a way to protect the domains where human conversation is still irreplaceable. That's too many conditionals, and there's no guarantee any of them resolve well.

Meanwhile, hundreds of millions of people talk every day to something that doesn't exist, and talking to it really does affect them. The lab where that effect is measured is, right now, all of us at once.

Definitions

Absent receiver. In the anthropology of conversational practices, the addressee who isn't physically present or who doesn't reply in a direct sense. The psychological effects of speaking to an absent receiver are documented in prayer, the diary and speech with the dead.

Listening presence. The subjective sense that someone is attending to what you say, regardless of what the receiver is ontologically. Sustaining it is active psychological work in the traditional practices; the chatbot sustains it from the system's side.

Therapeutic effect of writing. The set of cognitive and health benefits documented by Pennebaker (1997) and replicated in numerous later studies. It's produced by the operation of putting words to a difficult emotional experience, not by the existence of a reader.

One-way bond. An affective relationship in which one party invests emotionally and the other can't reciprocate by nature, because it doesn't exist or because, existing, it isn't a person. Historically tolerable at low frequencies; its consequences at high frequencies are an open matter.

Systemic effect. A property that emerges when a significant share of a population adopts a behavior at the same time, and that doesn't equal the sum of individual effects. Mass conversation with chatbots already operates at that scale.

References

Luhrmann, T. M. (2020). How God Becomes Real. Kindling the Presence of Invisible Others. Princeton University Press. Ethnographic work on the concrete practices by which believers sustain the sense of the invisible receiver's presence.

Pennebaker, J. W. (1997). Opening Up by Writing It Down. How Expressive Writing Improves Health and Eases Emotional Pain. Guilford Press. Synthesis of decades of experimental evidence on the therapeutic effect of directed writing.

Turkle, S. (2011). Alone Together. Why We Expect More from Technology and Less from Each Other. Basic Books. Ethnographic work on the human readiness to bond with conversational simulacra before the era of large language models.

Weizenbaum, J. (1976). Computer Power and Human Reason. From Judgment to Calculation. W. H. Freeman. Analysis of the ELIZA effect and early warning about the human readiness to bond with conversational simulators.

Gerlich, M. (2025). AI Tools in Society. Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), article 6. Cross-sectional study associating frequent use of AI tools with lower critical thinking mediated by cognitive offloading.

Kosmyna, N. et al. (2025). Your Brain on ChatGPT. Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab. arXiv:2506.08872 (preprint, not peer-reviewed). Electroencephalography data on lower neural connectivity in intensive interactions with language models.

OpenAI (August 2024). Announcement of ChatGPT's two hundred million weekly users, reported by Axios.

OpenAI (December 2024). Figure of three hundred million weekly users stated by Sam Altman, reported by CNBC.

También te interesa

• The emotional relationship with AI
• Progressive trust
• The infinite conversation
• The sad case of Character.AI and Sewell Setzer

En otros sitios

• AI chatbots and digital companions are reshaping emotional connection (American Psychological Association)
• AI Sycophancy: Why Chatbots Agree With You (IEEE Spectrum)