Why Do People Believe a Chatbot Has Feelings? Part One: What People Actually Think

Abstract

Our other series in this department asks whether an AI system could be conscious. This one asks a different question, with a firmer answer available: what do people already believe, and why?

A 2024 study of a nationally representative sample of 300 US adults found that two-thirds attributed at least some possibility of subjective experience to ChatGPT.

The finding that matters most is not the headline. It is that the belief was strongest among the people who used it most — which is the opposite of what familiarity is supposed to do.

1.Why This Is Not a Trivia Question

Clara Colombatto, at University College London and the University of Waterloo, and Stephen Fleming at UCL open by giving two reasons why folk attributions of consciousness to AI are worth measuring.

The first is practical. Attributions of consciousness may mediate future moral concern towards AI regardless of whether the systems are actually conscious. What people believe about machine minds will shape regulation, product design and public argument long before the underlying question is settled — if it ever is.

The second is scientific. How ordinary people decide that something has an inner life is itself a finding about human cognition, and machines are a genuinely new test case: highly articulate, entirely non-biological, and available to everyone.

Note what this study is not. It measures beliefs about consciousness, not consciousness. Nothing here bears on whether ChatGPT has experiences, and the authors do not suggest otherwise.

2.What They Did

They recruited a nationally representative sample of 300 US adults through Prolific, stratified to match population characteristics — 152 women, 142 men, two non-binary, four not answering, mean age 46.

Participants rated how capable they thought ChatGPT was of having subjective experience, on a scale adapted from earlier work on attributions of phenomenality. They were also asked about mental states typically associated with consciousness, and about how often they used the system.

Sample size and representativeness are the strengths here. Three hundred people is small for a survey but adequate for the effects reported, and national stratification means this is not a convenience sample of students or technology enthusiasts.

3.What People Said

A third of participants — 33 per cent — reported that ChatGPT was definitely not an experiencer. The remaining 67 per cent attributed some possibility of phenomenal consciousness.

The distribution matters more than the split. The mean rating was about 25 on the scale used, and the median was 16 — well below the midpoint. So the modal position is not that ChatGPT is conscious. It is that it might be, a little, and one cannot be certain.

That is a more defensible position than either confident answer, and it is worth saying so plainly rather than treating the 67 per cent as evidence of public credulity.

The attributions also behaved coherently. They predicted attributions of specific mental states associated with phenomenality, which means participants were not simply agreeing with an odd question — they held a consistent underlying view.

4.The Result That Inverts the Expectation

Attributions increased with usage frequency. In the authors' summary: those who interact with ChatGPT more frequently are also more likely to believe that it has subjective experiences.

The intuitive expectation runs the other way. Familiarity with a technology is generally assumed to demystify it — the more you use the thing, the more you understand it as a mechanism, and the less you are inclined to see a mind in it.

That is not what happens here. Whatever using a language model repeatedly does to a person, it does not reliably reduce the sense that something is on the other side.

Two readings are available and this study cannot separate them. Use may cause the belief: the accumulated experience of coherent, responsive conversation supplies something that reads as evidence. Or the belief may cause the use: people already inclined to see a mind there talk to it more.

This is a cross-sectional survey. The direction of causation is not established, and the authors do not claim it is.

5.What the Study Cannot Settle

US adults only, one system, one moment in 2024. Attitudes to a technology moving this fast have a short half-life, and a study of ChatGPT in 2024 may not describe attitudes to the systems people use now.

The authors also note a limitation worth carrying forward: they measured usage frequency, not understanding. Whether knowing how these systems work technically changes the attribution is a separate question and an open one.

That gap is where this series goes next. Knowing that a belief tracks experience is one thing; understanding the mechanism by which experience produces belief is another, and it turns out there is a well-developed literature on exactly that — developed for an entirely different subject.

Editorial Comment

MindHeaven® sells supplements. We have no chatbot, no companion product, and no position on machine consciousness to protect.

We are covering this because the finding is uncomfortable in a way that generalises past its subject. Repeated first-hand experience of something is normally treated as the best kind of evidence a person can have — better than being told, better than reading about it. Here it is associated with a belief that the best available scientific assessment, covered in our other series, does not support.

That should give pause to anyone in our business. "I felt it work" is the most persuasive sentence in the supplement industry, and it is the same class of evidence.

Part Two asks how experience generates conviction, using a framework built to explain belief in spirits and flat earths rather than in machines.