I’ve been arguing for a while that AI literacy isn’t just about knowing what AI can do. It’s also about understanding what students think AI is. The mental models students bring into the classroom shape how they use AI, what they trust it for, and where they question it.
A relatively recent study by Colombatto, Birch, and Fleming (2025) in Communications Psychology puts a sharper edge on this argument. Their picture of AI trust adds layers that the standard “anthropomorphism increases trust” story doesn’t cover.
The authors ran a pre-registered experiment with 410 US adults. Participants rated ChatGPT on capacity for consciousness and 20 specific mental states. They also completed a general knowledge task (country populations) where they could revise their answers based on what they were told was “ChatGPT’s advice.” The advice was actually pre-generated based on prior human performance, which let the researchers control accuracy without participants knowing.
The behavioral measure was simple: how often did participants revise their initial answer to match the advice? The authors call this “advice-taking” and use it as a behavioral proxy for trust.
Fifty-seven percent of participants attributed some possibility of consciousness to ChatGPT. This is striking on its own. Despite expert opinion strongly disagreeing, a majority of US adults think there might be something it is like to be ChatGPT.
Bayesian analyses produced strong evidence against a positive correlation between consciousness attribution and advice-taking. People who saw ChatGPT as more conscious did not trust its advice more. If anything, the relationship trended slightly negative.
The authors put it plainly: “Despite common concerns about the consequences of consciousness attributions to AI for user trust, we found strong evidence against a positive association between consciousness attribution and advice-taking.” (p. 5)
Smart AI vs Emotional AI: Two Dimensions of Mental State Attribution
The conceptual move that makes the paper useful is the recognition that mental state attribution isn’t one thing. Social psychology has long separated two factors: intelligence (reasoning, memory, knowing, planning) and experience (emotions, feelings, sensations). Colombatto and colleagues asked whether these two dimensions affect AI trust differently.
The answer is yes. Attributions of intelligence were strongly positively correlated with advice-taking. Attributions of experience were weakly negatively correlated with advice-taking. The two dimensions pulled in opposite directions. Consciousness attributions, which blend both, washed out to null or slightly negative.
The proposed mechanism is intuitive. The authors suggest that “the negative relationship between experience-related attributions and advice acceptance may stem from an interpretation of emotional AI as more volatile, biased, or unpredictable, leading to scepticism about its ability to provide accurate advice on factual tasks” (p. 5). Emotional AI feels less reliable for facts. Smart AI feels more reliable for facts.
Say-Do Gap on Trust
A second finding deserves attention. Consciousness attributions were positively correlated with self-reported trust. They were not correlated with actual advice-taking behavior. People said they trusted ChatGPT more when they saw it as conscious. They didn’t act on that trust.
This gap between what users say and what they do connects to Liu et al.’s (2026) work on AI assistance and persistence, where students’ self-reported AI experiences didn’t fully predict their behavioral patterns. Self-report alone is not enough for understanding how students actually use AI.
AI Trust in the Classroom
The dimension finding matters for AI literacy. When we shape how students think about AI, we shape how they trust it. When teachers describe AI as “knowing” or “smart,” student trust goes up. Words like “feeling” or “emotional” have the opposite effect on factual tasks. Both framings are misleading in their own ways, since neither matches what LLMs actually do.
Shanahan (2024) argued in his philosophical analysis of LLM language that loose use of words like “thinks” and “knows” obfuscates mechanism. This new study gives that argument empirical weight. Those words don’t just describe systems imprecisely. They shift user behavior. The metaphors we use for AI, as Roe, Furze, and Perkins (2025) argue in their “digital plastic” work, shape how students engage with these tools.
For u teachers and educators, this means two things. First, the language we model when we talk about AI in class carries weight. Second, what looks like critical engagement (students saying they don’t fully trust AI) might mask high behavioral reliance. The say-do gap is real.
Limitations
The advice task was factual. The authors flag that experience attributions might matter more on emotional tasks like relationship advice or mental health support. A student turning to AI for emotional reasons probably draws on a different set of attributions than a student fact-checking a history claim. The two-dimension framework is useful but the relative weight likely shifts by task type.
The authors conclude that “these findings highlight how human–AI interactions are shaped by complex and multifaceted inferences about the capacities of AI systems” (p. 6). That’s the framing teachers can take into the classroom. Students aren’t approaching AI with a single belief. They’re approaching it with a layered set of attributions that shape both what they say and what they do. AI literacy education needs to engage with both layers.
References
- Colombatto, C., Birch, J., & Fleming, S. M. (2025). The influence of mental state attributions on trust in large language models. Communications Psychology, 3(1), 84. https://doi.org/10.1038/s44271-025-00262-1
- Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., & Dubey, R. (2026). AI assistance reduces persistence and hurts independent performance [Preprint]. arXiv. https://arxiv.org/abs/2604.04721.
- Roe, J., Furze, L., & Perkins, M. (2025). Digital plastic: A metaphorical framework for Critical AI Literacy in the multiliteracies era. Pedagogies: An International Journal. Advance online publication. https://doi.org/10.1080/1554480X.2025.2557491
- Shanahan, M. (2024). Talking about large language models. Communications of the ACM, 67(2), 68-79. https://doi.org/10.1145/3624724.