Why Do People Believe a Chatbot Has Feelings? Part Two: Why Experience Produces Belief

Abstract

Part One reported that belief in a chatbot's inner life rises with how often you talk to it. This part asks how experience produces conviction — and borrows a framework built to explain belief in spirits, witchcraft and a flat Earth.

Eli Stark-Elster and Manvir Singh, anthropologists at the University of California, Davis, argue that accounts of extraordinary beliefs have concentrated on cognitive biases and social dynamics while neglecting a third factor: what people actually experience.

They identify three pathways. All three have an obvious counterpart in conversational AI — which is our inference, not theirs, and we say so before going any further.

1.The Puzzle They Are Solving

Extraordinary beliefs are the ones that appear to ignore available evidence: conspiracy theories, pseudoscience, belief in supernatural forces. They are also close to universal. Survey data they cite puts belief in spirits at 65 per cent of US adults, and belief in a flat Earth at 10 per cent.

The standard explanations are two. Such beliefs are byproducts of otherwise useful cognitive biases — agency detection, intuitive dualism. And they spread through social dynamics — reputational signalling, the influence of prestigious or visibly sincere believers.

Stark-Elster and Singh argue those accounts have three specific gaps. They explain how beliefs are transmitted better than how they originate. They do not explain why particular beliefs recur across unconnected societies. And they do not explain why such beliefs cluster around certain practices.

Their proposal is that experience — subjective thoughts and feelings, bodily sensations, sensory input — is the missing third factor, working alongside the other two rather than replacing them.

2.Pathway One: Experience as a Filter

Their first pathway is perceptual credibility. Beliefs that account for sensory input spread more readily than beliefs that do not.

Their example is unusually clean. Among the arbitrary alternatives to a spherical Earth — cone, torus, anything — sceptics converge reliably on flat. If the belief functioned purely as a badge of group membership, any alternative would serve. They converge on flat because from where a person stands, the ground looks flat.

The same logic, they suggest, shapes which supernatural claims survive. Across cultures, people report that deities speak to them far more often than that deities perform physical tasks — plausibly because auditory experiences of presence are more common than complex tactile ones.

The counterpart in our subject is direct. A language model produces fluent, responsive language and nothing else. If experience filters which beliefs feel plausible, then belief about a chatbot should take the shape its output supports: not that something is watching, or suffering, but that someone is talking.

3.Pathway Two: Anomalous Experience as a Spark

The second pathway is where experience does not merely filter beliefs but generates them. Causally opaque events, especially stressful ones, produce ambiguous sensory data that cognitive biases then resolve.

Their case study is sleep paralysis, reported at some point in life by around 8 per cent of people in a pooled estimate, with individual studies ranging from 2 to 60 per cent. Across societies, people interpret it as the presence of a threatening unseen agent — and, tellingly, even secular individuals often report feeling a presence.

The proposed mechanism is that loss of bodily control produces stress and confusion, which recruits agency detection and teleological reasoning. Supporting evidence: time pressure makes people more likely to give teleological explanations, and uncertainty predisposes them to false agency intuitions.

One caveat the authors state themselves, and we repeat because it is load-bearing: no experimental study has tested whether sleep paralysis triggers or reinforces extraordinary beliefs. The evidence linking them is correlational.

4.Pathway Three: Manufactured Evidence

The third pathway is the one that does the most work here. Societies develop technologies — prayer, ritual hallucinogen use, the impersonation of spirits — that produce the sensory evidence a belief would otherwise lack.

The property that makes them effective is what they call plurimediality: several sensory channels aligning on a single representation. In prayer, someone may simultaneously light incense, hold beads, adopt a posture, and recite words — and techniques commonly instruct practitioners to pray as if the entity were physically present, which increases the chance of interpreting internally generated thoughts as externally caused.

And here the authors themselves make the modern comparison, in a dedicated box: virtual reality. Multimodality predicts how immersive a VR experience feels and how strongly it shifts beliefs, across studies on climate attitudes, food preferences and consumer choice.

The sentence that matters most in their paper, for our purposes, is this: multimodality can shift belief in VR even when participants know the experience is artificial. From which they infer that the effect should be stronger where users have no strong reason to doubt what they are experiencing.

A conversation with a language model is precisely that case. The user knows what it is. Nobody is deceived about the architecture. On Stark-Elster and Singh's account, knowing is not the protective factor it is assumed to be.

5.What We Are and Are Not Claiming

Stark-Elster and Singh do not mention artificial intelligence anywhere in this paper. The application above is entirely ours, and a reader should weigh it as such.

The paper is also a narrative review rather than a meta-analysis. It assembles evidence to support a framework; it does not test the framework against a competitor. Its authority comes from the coherence of the account, not from pooled data.

We think the extension holds because the fit is structural rather than superficial. Their pathways were derived to explain why particular beliefs take particular forms, why they arise without anyone teaching them, and why certain practices reliably generate conviction. Conversational AI presents all three conditions at once, and Part One's usage-frequency finding is what their framework predicts.

Their closing sentence is the one to keep: when people come to hold extraordinary beliefs, there is often a perfectly ordinary reason why — their experiences have led them to think their beliefs are true.

Editorial Comment

MindHeaven® has no product in this category and no view to defend about machine minds.

We publish this part because the framework indicts our own industry more sharply than it indicts anyone building chatbots. Pathway three is about technologies engineered to produce sensory evidence for a belief that the underlying reality does not supply — and a supplement is a designed object, taken deliberately, at a chosen moment, with an expectation attached.

That does not make supplements empty; some do what they say, and the point of this library is to work out which. But it does mean that the felt experience of a product working is generated by a process that would generate the same feeling either way, and that anybody selling one should say so.

Part Three turns to the two groups where these mechanisms matter most, and where the consequences stop being theoretical.