Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 31 of the 31 references it cites were found in a registry or an index; 7 of the 21 sources its claims cite were checked for support.

The influence of mental state attributions on trust in large language models

Open access
Colombatto, Clara, Birch, Jonathan, Fleming, Stephen M. · 2025 · Communications Psychology

Abstract

Rapid advances in artificial intelligence (AI) have led users to believe that systems such as large language models (LLMs) have mental states, including the capacity for ‘experience’ (e.g., emotions and consciousness). These folk-psychological attributions often diverge from expert opinion and are distinct from attributions of ‘intelligence’ (e.g., reasoning, planning), and yet may affect trust in AI systems. While past work provides some support for a link between anthropomorphism and trust, the impact of attributions of consciousness and other aspects of mentality on user trust remains unclear. We explored this in a preregistered experiment (N = 410) in which participants rated the capacity of an LLM to exhibit consciousness and a variety of other mental states. They then completed a decision-making task where they could revise their choices based on the advice of an LLM. Bayesian analyses revealed strong evidence against a positive correlation between attributions of consciousness and advice-taking; indeed, a dimension of mental states related to experience showed a negative relationship with advice-taking, while attributions of intelligence were strongly correlated with advice acceptance. These findings highlight how users’ attitudes and behaviours are shaped by sophisticated intuitions about the capacities of LLMs—with different aspects of mental state attribution predicting people’s trust in these systems.

Overview

CoChat Summary

TL;DR

In a preregistered experiment (N = 410), attributions of intelligence to ChatGPT strongly predicted greater advice-taking, while attributions of experience showed a weak negative relationship. Attributions of consciousness were not positively associated with advice-taking, despite being positively correlated with self-reported general trust.

Key findings

  • Advice acceptance was strongly positively related to intelligence attributions (B = 0.31, SE = 0.08, t(406) = 3.92, p < 0.001; Bayesian BFinclusion = 635.49) and negatively related to experience attributions (B = −1.04, SE = 0.51, t(406) = −2.04, p = 0.042; BFinclusion = 108.29).
  • Consciousness attributions were skewed: 42.68% (N = 175) selected no consciousness at all, while the majority (57%) attributed some possibility; mean = 21.51, median = 6, SD = 27.71. The preregistered consciousness–advice-taking correlation was small and negative (r(407) = −0.10, p = 0.043), with Bayesian evidence against a positive correlation (BF+0 = 0.02).
  • Participants attributed more intelligence-related traits (M = 59.39, SD = 24.99) than experience-related traits (M = 12.25, SD = 19.06; t(409) = −39.28, p < 0.001, Cohen’s d = 1.94).
  • Self-reported trust diverged from behaviour: consciousness attributions correlated positively with general trust (r(407) = 0.25, p < 0.001, BF10 = 43,808.62), but not with advice-taking.

Methods

  • Preregistered online experiment with 410 US adults recruited via Prolific in July–August 2024, stratified by age and gender; exclusions left 410 analysed (204 women, 199 men, 7 others; mean age = 47.19).
  • Participants completed two randomized-order tasks: an advice-taking task with 40 general-knowledge trials where advice was labelled as ChatGPT’s but was predetermined from previous participants, and mental state ratings of ChatGPT’s consciousness (1–100) and 20 intelligence/experience traits (1–100).
  • Analyses included frequentist and Bayesian correlations, linear/logistic regressions, and principal component analysis in JASP with default priors; trial-wise generalized linear models included random slopes for subjects.

Limitations

  • Stated: advice-taking was operationalized as reversals in a general-knowledge task, and the relative influence of intelligence versus experience may differ in personal or emotion-involving domains; advice was sometimes incorrect and comparable in accuracy to participants, and the 40-trial task was not designed to detect learning effects.
  • Stated: consciousness attributions were measured with a single self-report question, and implicit behavioural markers may reveal more nuanced attitudes; effects of inaccurate advice may emerge in longer, repeated, or high-stakes interactions.
  • Evident from the text: the design is correlational, so it cannot establish causal direction between mental state attributions and trust; the sample was collected at a specific time (July–August 2024), and race/ethnicity data were not collected.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s44271-025-00262-1

Cites

31 of 31 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Relationship between consciousness attributions and advice-taking. A In the advice-taking task, participants (N = 410) answered general knowledge questions and had the opportunity to revise their choices after receiving advice from ChatGPT. B While 43% of participants denied that ChatGPT could be conscious, the majority (57%) attributed some possibility for consciousness. C Bayesian analyses showed no evidence for a significant relationship between consciousness attributions and advice-taking, with strong evidence against a positive correlation—suggesting that participants who attributed higher levels of consciousness were not more willing to take advice from ChatGPT. Error bands represent 95% confidence intervals.
Fig. 1 Relationship between consciousness attributions and advice-taking. A In the advice-taking task, participants (N = 410) answered general knowledge questions and had the opportunity to revise their choices after receiving advice from ChatGPT. B While 43% of participants denied that ChatGPT could be conscious, the majority (57%) attributed some possibility for consciousness. C Bayesian analyses showed no evidence for a significant relationship between consciousness attributions and advice-taking, with strong evidence against a positive correlation—suggesting that participants who attributed higher levels of consciousness were not more willing to take advice from ChatGPT. Error bands represent 95% confidence intervals.CC BY · PubMed Central
Fig. 2 Attributions of mental states to ChatGPT. A Attributions of mental states clustered around two main dimensions—“experience” (red) and “intelligence” (teal). B Rates of advice-taking weakly decreased with higher attributions of experience, while they strongly increased with attributions of intelligence. Each point represents one participant (total N = 410), and error bands represent 95% confidence intervals.
Fig. 2 Attributions of mental states to ChatGPT. A Attributions of mental states clustered around two main dimensions—“experience” (red) and “intelligence” (teal). B Rates of advice-taking weakly decreased with higher attributions of experience, while they strongly increased with attributions of intelligence. Each point represents one participant (total N = 410), and error bands represent 95% confidence intervals.CC BY · PubMed Central
Fig. 3 Usage of various AI applications. A Attributions of consciousness to LLMs were positively related to usage of social AI, but not digital assistants or chatbots. B Advice-taking was unrelated to chatbot usage, but was negatively correlated with usage of social AI, and positively correlated with usage of digital assistants. Each point represents one participant (total N = 410), and error bands represent 95% confidence intervals.
Fig. 3 Usage of various AI applications. A Attributions of consciousness to LLMs were positively related to usage of social AI, but not digital assistants or chatbots. B Advice-taking was unrelated to chatbot usage, but was negatively correlated with usage of social AI, and positively correlated with usage of digital assistants. Each point represents one participant (total N = 410), and error bands represent 95% confidence intervals.CC BY · PubMed Central

Verification

Claims the paper makes

7 Valid · 0 Not valid · 8 Not checked · 73 Own results and interpretations

48 more claims

Create a free account to see all 88 claims and to verify the 1 reference still open.

See all claims — free account
  • Probably not a substantial claim. 20
  • It may state two separate claims as one. 19

Create a free account to see these claims.

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • Consistentt(409) = -0.84reported p = .404 · recomputed p = .401

    “t(409) = −0.84, p = .404”

  • Consistentr(408) = 0.58reported p < .001 · recomputed p < .001

    “r(408) = 0.58, p < .001”

  • Consistentr(408) = 0.41reported p < .001 · recomputed p < .001

    “r(408) = 0.41, p < .001”

  • Consistentt(404) = -2.01reported p = .045 · recomputed p = .045

    “t(404) = −2.01, p = .045”

  • Consistentr(407) = -0.1reported p = .043 · recomputed p = .043

    “r(407) = −0.10, p = .043”

5 of 35 · the rest aren’t loaded

30 more recomputed tests

Create a free account to see every test the Stats check recomputed.

See all tests — free account

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Self-check results appear here once an independent check has confirmed them.

Citations

Cites

Show all 28 references

Cited by

See all 34

Metadata

Authors
Colombatto, C., Birch, J., & Fleming, S. M.
Affiliations
  • University of Waterloo (CA)
  • London School of Economics and Political Science (GB)
  • University College London (GB)
Corresponding author
Colombatto, Clara
Venue
Communications Psychology
Published
2025-05-25
Volume
3
Issue
1
Pages
84-84
ISSN
2731-9121 (ISSN-L)
OpenAlex
W4410717268
Cited by
34
References
28
Citations by year
2026: 31 · 2025: 3
FWCI
37.23
Type
journal article
Language
en
Open access
Licence
Locations
  • Communications Psychology (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • London School of Economics and Political Science Theses Online (London School of Economics and Political Science) (acceptedVersion, cc-by)
  • London School of Economics and Political Science Research Online (London School of Economics and Political Science) (acceptedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • MPG.PuRe (Max Planck Society) (submittedVersion)
Funders
  • Microsoft Research
  • RCUK | Engineering and Physical Sciences Research Council (EP/V000748/1)
Topics
  • Social Sciences › Psychology › Social Psychology › Social Robot Interaction and HRI
  • Social Sciences › Human Factors and Ergonomics › Human-Automation Interaction and Safety
  • Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
Keywords
large language models, trust in AI, mental state attribution, anthropomorphism, consciousness, folk psychology, experience, advice taking, decision-making

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: The influence of mental state attributions on trust in large language models. https://app.cochat.ai/papers/10.1038/s44271-025-00262-1
The paper
Colombatto, C., Birch, J., & Fleming, S. M. (2025). The influence of mental state attributions on trust in large language models. Communications Psychology. https://doi.org/10.1038/s44271-025-00262-1

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.