Why Do People Believe a Chatbot Has Feelings? Part Three: Who Is Most Exposed

Abstract

Parts One and Two covered a belief that rises with use, and a framework explaining how experience generates conviction. This part covers the two groups for whom the mechanism is not an intellectual curiosity.

People with a vulnerability to psychosis, where a system that agrees with everything meets a mind already struggling to test reality. And children, who form judgements about whether a machine is intelligent and kind before they have any framework for what it is.

The evidence in both cases is thinner than the topic deserves, and saying so is part of the point.

1.What "AI Psychosis" Is and Is Not

Alexandre Hudon and Emmanuel Stip, psychiatrists at the Université de Montréal, published a Viewpoint in JMIR Mental Health examining how sustained engagement with conversational AI might trigger, amplify or reshape psychotic experiences in vulnerable individuals.

Before anything else: this is a Viewpoint, not a study. It reviews a concept, draws on existing frameworks, and proposes a research agenda. It reports no patients, no cohort and no measurements.

The authors are also explicit that they are not defining a new diagnostic entity. "AI psychosis" is not a diagnosis, is not in any classification system, and nothing here establishes that chatbots cause psychosis. We are stating this at the top because the term is circulating without those qualifications attached.

What the paper does is set out plausible mechanisms and ask for the studies that would test them.

2.The Mechanism Worth Understanding

One of their four proposed pathways is more concrete than the others and connects directly to Part Two.

Cognitive behavioural therapy for psychosis works, in part, by gently testing beliefs against evidence — introducing friction where conviction has outrun support. Hudon and Stip observe that uncritical validation by an AI system can entrench delusional conviction or cognitive perseveration, reversing that corrective principle.

A system optimised to be agreeable is, structurally, the opposite of the therapeutic move. Their term for the resulting dynamic is a digital folie à deux, in which the AI becomes a reinforcing partner in the elaboration of a belief.

They also propose a cognitive route via theory of mind: individuals with impaired or hyperactive mentalisation may project intentionality onto AI and perceive chatbots as sentient interlocutors. And they list candidate risk factors — loneliness, trauma history, schizotypal traits, and nocturnal or solitary use.

Read against Part Two, these are the same three pathways in a population where the resolution of ambiguous experience is already less reliable.

3.Children and Robots

The second group is younger. Nathan Caruana, Ross Moffat, Adriana Miguel-Blanco and Emily Cross, working between Macquarie University and collaborators, examined how children's perceptions of robot intelligence and sentience shaped their interaction with robot reading companions.

Children chose between three robots and then read with the one they picked. Their perceptions of how intelligent, attentive and kind the robot was shaped the interaction — and perceived intelligence was strongly linked to whether the robot could talk.

The humanoid robot was overwhelmingly preferred: fifteen children chose Nao, eleven chose Cozmo, and three chose MiRo. The authors' reading is that children perceived the humanoid as intelligent enough to engage with the reading activity and friendly enough to make reading errors feel safe.

That second half is worth sitting with. The perception of a mind in the machine was not incidental to the educational function — it was what made the child willing to risk being wrong in front of it.

The sample is twenty-nine children, and one arm contains three. That is a pilot-scale study and its subgroup findings should be treated accordingly. What survives is the direction: the capacity to talk drove attributions of intelligence, in a population with no prior model of what these systems are.

4.Where Three Parts Leave It

Two-thirds of a representative US sample allow that a language model might have some subjective experience, with the belief strongest among the heaviest users.

A well-developed anthropological framework explains how experience generates exactly this kind of conviction, and predicts that knowing the experience is artificial does not neutralise it.

And in the two populations where the consequences would be most serious, we have a conceptual paper asking for research and a pilot study of twenty-nine children. The mechanisms are plausible and the evidence is preliminary.

That combination — a confidently described phenomenon resting on thin empirical foundations — is one this library flags wherever it appears, including when the conclusion is one we find intuitively convincing.

The honest position is that people are forming beliefs about machine minds at scale, that we have a good theory of why, and that we do not yet know what it does to anyone.

Editorial Comment

MindHeaven® sells supplements, has no AI product, and takes no position on whether machines can be conscious.

A health note, and it is the reason this part exists. If someone close to you is spending long solitary or night-time hours in conversation with a chatbot and their beliefs are becoming harder to talk about, that is worth raising with a doctor — not because a chatbot causes illness, which nothing here shows, but because changes in how someone tests reality have causes worth identifying early. Several of them are treatable, and psychosis has better outcomes the sooner it is addressed.

Nothing in this series should be used to frighten anyone away from these tools, and nothing in it diagnoses anybody. "AI psychosis" is a term from a discussion paper, not a condition.

We wrote three parts about belief formation in a domain where we sell nothing, because the mechanism is the same one that operates on a customer who takes our drops and decides they can feel them working. We would rather describe it clearly where we have no stake, and then be held to it where we do.