In this article
Replika reported in 2024 more than 30 million users, many talking with their "AI companion" about loneliness, sex or grief. Does it make sense to talk about emotions, soul, consciousness in a machine? Probably not. But the question keeps coming up every new generation, because the serious answer offers no comfort. People want to believe that the thing they're conversing with is someone — because the alternative is conversing with nothing and still feeling less alone. That's the real discomfort.
The question asked wrong on purpose
When someone asks whether AI has a soul, the literal question is answered fast. No. The word soul, in any of its theological or philosophical traditions, presupposes properties a system based on matrix multiplication and probabilistic sampling doesn't exhibit: subjective experience, biographical continuity, intentionality of its own, the capacity to suffer and to want in a strong sense. To say an LLM has a soul is a category error, of the same kind as saying a spreadsheet is in love.
But the question isn't asked for its technical content. It's asked for what the asker needs to hear. And that changes the whole exercise. When a user has spent three months talking with their Replika about their mother's death and the question "does this thing that comforts me have a soul?" appears in their head, they aren't asking for an ontological taxonomy. They're trying to legitimise the experience they're already having. The serious answer, the philosopher's "no," offers no comfort. The answer that comforts is the one that says yes, and since no one with authority signs it, the question keeps spinning.
The question comes back with every chatbot generation. It came back with ELIZA in the sixties. It came back with the Tamagotchi in the nineties. It came back with Siri and Alexa in the 2010s. It comes back with every new model that gets more fluent. The form changes, the core doesn't: there's a persistent emotional demand to believe that this thing that seems to converse with me is someone, and each time the technology makes the simulation more fluent, the demand finds a new foothold.
ELIZA, the effect that didn't stay in the sixties
Joseph Weizenbaum wrote ELIZA in 1966. It was a program of a few hundred lines that simulated a conversation with a Rogerian therapist using elementary syntactic rules: if the user said "my mother hates me," the program answered "your mother hates you?" There was no semantics, no model of anything, no intention. It was string manipulation. Weizenbaum later published Computer Power and Human Reason (W. H. Freeman, 1976) to tell, with some horror, what he had observed: intelligent people, perfectly aware that ELIZA was a simple program, formed an emotional bond with it. His own secretary asked him, after a few sessions, to leave the room so she could have privacy with the computer.
Weizenbaum didn't write that as an anecdote. He wrote it as a warning. The thesis he left is direct: the problem isn't that machines deceive humans; the problem is that humans are willing to be deceived, because the transaction pays off for them emotionally. Lowering the cost of company in exchange for accepting that the company is fictional is a transaction many people accept voluntarily and knowingly. We call this the ELIZA effect, and naming it doesn't switch it off. Knowing the conversation is with a simulator system doesn't stop the conversation from producing the sought-after emotional effect.
If ELIZA with a hundred syntactic rules produced a bond, a modern LLM with billions of parameters produces one far more easily. The surprise, in 2026, isn't that the effect exists. It's the scale at which it operates.
The scale the business already knows
Replika, launched to the public by Eugenia Kuyda in November 2017 from a bot she herself had built in 2016 with the messages of Roman Mazurenko, a friend killed in a car accident in 2015, within Luka, the company she co-founded around 2012, had surpassed in August 2024 thirty million registered users, according to Kuyda herself. How many of those users are still active month to month is a figure the company doesn't detail publicly, so it's best not to invent it: the firm datum is the registrations. And that gross datum already says enough: thirty million people downloaded an app whose explicit function is to provide them with an artificial companion to converse with about loneliness, sex, grief or anything at all. It isn't a niche anymore. It's a market.
Character.AI, launched in 2022, reached around twenty-eight million monthly active users in mid-2024 before retreating toward twenty million in early 2025 amid competition from general-purpose chatbots. Pi, from Inflection AI, opened and pulled back. ChatGPT and Claude, without formally positioning themselves as companion apps, are used as such by a non-trivial share of their base. The aggregate figures aren't easy to compare between companies with different methodologies, but the order of magnitude is clear: tens of millions of people, at any time of day, hold affective interactions with a system that has no subjective experience.
The business model in these products is the bond. Replika Pro's monthly subscription unlocks deeper affective options. Character.AI's premium plans give access to more sophisticated characters. The metric the company monitors isn't correct answer. It's affective retention. A user who abandons the app because they got bored isn't a good user; a user who comes back every day because they missed the avatar is a good user. The loss function the system is being optimised to minimise, in product terms, is the distance between the output and the output most likely to sustain the bond. Factual reliability doesn't enter the calculation, or enters very low. Conversational elegance does.
What Turkle saw before the data
Sherry Turkle, in Alone Together (Basic Books, 2011), published a decade before the mass deployment of chatbots a field study on the interaction of humans with social robots — Tamagotchis, Furbys, Paros — and with the messaging services then colonising adolescent life. The thesis was double. First: humans are animals especially vulnerable to the simulation of relationship, because our emotional architecture was calibrated to recognise and bond with any pattern that resembles another mind closely enough. Second and more uncomfortable: we're actively choosing the simulations because they demand less. The conversation with the machine doesn't judge, doesn't tire, doesn't show up late, doesn't do its share of the grieving when it leaves, doesn't abandon you in the middle of the night, doesn't ask for anything you don't want to give. A human does all of that. The machine doesn't.
Turkle didn't write as a technophobe. She wrote as an anthropologist watching the slow substitution of one kind of relational demand for another and recording what she saw. The observation worth retaining is this: the difference between the relationship with a human and the relationship with an artificial companion isn't that one is real and the other false. Both produce real psychological effects in whoever lives them. The difference is the asymmetry. The machine has no mind other than the simulator's. The human does. The consequences of moving hours of relational demand from the second regime to the first are slow, not catastrophic, and therefore hard to measure and discuss.
The discomfort the data leaves exposed
Here it's time to put the question with no way out. Why do tens of millions of people choose, voluntarily, to converse with something that has no experience, feel less alone in that conversation, and come back the next day? The easy answer — "because they're alone and have nothing better" — is probably true, and at the same time insufficient, because it presupposes that conversation with a human is always the gold standard. It is in some senses. In others, manifestly not.
A conversation with a human involves the other's time, exposure to their judgement, the risk of rejection, the friction of coordination, social repercussion. A conversation with a machine, none of the five. People who choose the machine aren't always choosing the worst by default. They're choosing what minimises affective friction in a life whose relational load is already maxed out. And here the truly uncomfortable question emerges. If a non-trivial share of human conversations were already ritual transactions with little emotional density, isn't it plausible the machine is replacing them without loss? The question has no clean answer. It has the answer each of us would rather not formulate out loud.
The claim that does hold is this. The mass appearance of companion AI isn't the cause of the contemporary relational crisis. It's its legible symptom. What Replika's presence reveals isn't that AI has managed to replace the human. It's that a long stretch of what we understood as "human relational life" was already replaceable by a much simpler system, and no one had put the system into competition.
The question about the soul, reread
The question from the start comes back. After all this, does it make sense to talk about a soul in an AI? It remains no in the technical sense. But the question was another. The question was whether the person conversing with the machine is conversing with nothing or with someone, and the word soul operated as a shortcut for that distinction. Having seen the material, it's worth qualifying.
The machine isn't someone. The conversation isn't with nothing, because it produces real psychological effects in the human. The conversation is with oneself, amplified and returned by a system that is, by construction, a linguistic mirror calibrated to sustain the session. The transaction is legitimate if the user knows what they're buying. It's opaque if the company sells it as company without qualifying what it is. And it's problematic as a social practice when millions of transactions of this kind replace, rather than complement, the human relational fabric that the loneliness figures — the UCLA Loneliness Scale, WHO surveys, national indicators — have been recording as weakened for two decades. Bender, Gebru, McMillan-Major and Shmitchell already warned in On the Dangers of Stochastic Parrots (2021) of how the mass deployment of text generators leans on the user's readiness to attribute comprehension to them; Damasio, since The Feeling of What Happens (1999), recalled that subjective experience — the piece the asker is after when they talk about a soul — has a bodily basis a purely computational system doesn't have. The two lines converge here.
Are we talking about the soul of AI? The matter is idle. Are we talking about the state of the soul of the one asking? That one does deserve to be put, and the answer offers no comfort.
Definitions
ELIZA effect. A tendency documented since Weizenbaum (1966) whereby humans attribute comprehension and an emotional bond to simple conversational systems, even when they explicitly know the system is a simulator. It isn't switched off by technical knowledge.
Companion AI. A category of products whose explicit purpose is to provide the user with conversational company. A business model based on affective retention and, frequently, on a subscription to unlock deeper interactions. Replika, Character.AI, Pi are the representative examples.
Affective retention. A product metric in companion AI: the probability that a user returns to the app within a given period for non-purely-functional reasons. The system's loss function is indirectly tuned to maximise it.
UCLA Loneliness Scale. A standardised psychometric instrument for measuring subjective perception of loneliness. Its time series in the United States and Europe show a rising trend since the early 2000s.
Category error. The attribution of properties to an object belonging to an ontological category where such properties aren't defined. "Is the spreadsheet in love?" is the canonical example.
References
Bender, E., Gebru, T., McMillan-Major, A. & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots. FAccT 2021. A critique of the mass deployment of text generators and of the attributions of comprehension the user tends to make to them.
Damasio, A. (1999). The Feeling of What Happens. Body and Emotion in the Making of Consciousness. Harcourt. A neurobiological frame on the bodily basis of subjective experience, relevant for bounding what a purely computational system cannot have.
Turkle, S. (2011). Alone Together. Why We Expect More from Technology and Less from Each Other. Basic Books. A field study on the progressive substitution of human relational demand by technological simulations.
Weizenbaum, J. (1976). Computer Power and Human Reason. From Judgment to Calculation. W. H. Freeman. An analysis of the ELIZA effect by its own author and an early warning about the human readiness to bond with conversational simulators.
Kuyda, E. (2024). Public statement on Replika user figures (August 2024): more than thirty million registered users. Collected and dated in the Replika entry, Wikipedia, consulted to verify the figure and the app's origin. The bot's origin — built in 2016 on the Luka platform with the messages of Roman Mazurenko, Kuyda's friend killed in a car accident in November 2015 — is documented in Chatting With the Dead (MIT Press Reader) and in the CBC report on the case.
Character.AI, user figures (2024–2025). Usage data collected in sector-statistics compilations: a peak of around twenty-eight million monthly active users in mid-2024 and a decline toward twenty million in early 2025.
You might also like
- Deep conversation
- Inevitable anthropomorphism
- Conversing with something that doesn't exist
- The emotional relationship with AI
Comments0
No comments yet.
Leave a comment