Why do people see ChatGPT so differently?

Two studies, three samples and a newly developed attitude scale

Study 1 began with 500 respondents from Serbia and retained 472 adults after excluding participants under 18, one person who did not consent to data processing, and gender categories that were too small for the planned moderation analysis. The final sample was 72% women, aged 18 to 58 with a mean age of 25.14. Participants completed Google Forms measures covering general attitudes towards AI, six HEXACO personality dimensions, right-wing authoritarianism, social dominance orientation, need for cognitive closure and four objective AI-literacy questions selected for lay users.

The researchers also built a 17-item scale through two pilot studies. It separated four ways of seeing ChatGPT: as a personal assistant, as a learning tool, as a risk of cognitive impairment and as a danger to humanity. That multidimensional design is useful because someone can value practical assistance and still worry about overreliance or wider harm. It also means the study measured respondents' beliefs about risks, not cognitive decline, learning performance, harmful behaviour or any observed effect of using ChatGPT.

Study 2 recruited 115 people in Germany and 119 in Serbia through Facebook, Instagram, LinkedIn, SurveySwap and Prolific. After exclusions, the paper's abstract reports 112 valid German and 117 valid Serbian respondents. Ages ranged from 18 to 71 in Germany and 18 to 61 in Serbia. The same survey framework was administered online, and the authors used group comparisons followed by exploratory path analysis. There was no representative probability sampling, behavioural experiment, follow-up period or product-use log.[1]

How large were the reported country-sample differences?

The strongest reported gap was on objective AI literacy: the German sample scored higher with a Cohen's d of 1.01, a large standardized difference. Positive attitudes towards AI were also higher in that sample, with d of 0.782. On the ChatGPT-specific scales, the German respondents scored higher for the learning-tool view, d 0.412, and personal-assistant view, d 0.301. The Serbian respondents scored higher for cognitive-impairment concern, d 0.749, and danger-to-humanity concern, d 0.346.

Those effect sizes describe separation in these recruited samples, not the share of either nation's population holding an opinion. The authors explicitly warn that they did not test measurement invariance—the requirement that a scale represents the same construct in the same way across groups. Without it, a mean difference can partly reflect translation, interpretation or response-style differences rather than only a difference in the underlying attitude. The samples were also small and self-selected, and the recruitment mix could differ by country in ways the model did not observe.[1]

AI literacy was associated with attitudes, not proven to change them

In the exploratory models, higher AI literacy was associated with more positive views in several places. In the Serbian Study 2 model, literacy related positively to seeing ChatGPT as a personal assistant and combined with openness to experience in some interactions. In the German model, higher literacy was associated with lower perceived danger, but it also had a small negative relationship with seeing ChatGPT as a learning tool. The mixed pattern is a warning against turning one correlation into the slogan that education automatically creates trust.

The survey measured literacy and attitudes at the same time. It cannot establish whether knowledge changed a person's judgement, whether favourable users learned more through experience, or whether education, occupation, income, media exposure and previous tool use affected both. A literacy programme could appropriately increase scepticism about some uses while increasing confidence in others. The goal for educators and public institutions should be calibrated judgement—understanding capabilities, failure modes and governance—not a predetermined rise in approval.[1]

Why the path models are hypothesis-generating

The researchers used path analysis to examine a network of proposed predictors and outcomes. They describe the Study 2 analysis as exploratory because it included data-informed modifications and the cross-country samples were small. An initial Serbian model had poor fit; after insignificant paths were removed, the revised model fit the same data better. The German model also produced acceptable fit indices, but fitting a modified model to one dataset does not show that the pattern will recur in a new group.

Some inputs were measured with short scales that had low internal-consistency estimates. In Study 2, HEXACO alpha values ranged from 0.31 to 0.56, social-dominance orientation was 0.37 in Serbia and 0.58 in Germany, and need for closure was 0.56 and 0.59. The authors report average inter-item correlations and explain why short scales can depress alpha, but low reliability can still make coefficients unstable. The danger-to-humanity scale reached 0.65 in the German sample. Individual paths, gender interactions and personality explanations therefore require independent replication, not policy built around a psychological profile.[1]

What the study means for public engagement and education

For schools, employers, newsrooms and governments, the practical message is that attitudes towards conversational AI are multidimensional and context-dependent. A person can accept a system as a convenient assistant while worrying about deskilling, misinformation or social power. Surveys that ask only whether people support or oppose AI will hide those distinctions. Engagement is more useful when it names a task, a benefit, a risk and the conditions under which use would be acceptable.

The results also argue for testing understanding rather than equating confidence or frequent use with literacy. The study used an objective knowledge measure, albeit only four selected questions. Public programmes could separately measure what people know, how they use tools and what outcomes follow, while offering accessible alternatives for people who choose not to use them. None of this requires labelling scepticism as ignorance or enthusiasm as competence. Both favourable and critical attitudes can coexist with accurate knowledge.

The gender analysis has an additional inclusion limit. Because very few respondents identified outside the man/woman categories, those respondents were excluded from moderation models, as were people who did not disclose gender in Study 1. The reported gender patterns therefore apply only to the analysed binary groups and do not provide evidence about non-binary or transgender people. Larger, deliberately inclusive samples would be needed to examine those experiences responsibly.[1]

What would change the assessment

A stronger cross-country test would preregister hypotheses, use representative or carefully stratified samples, establish translation quality and measurement invariance, and validate the path structure on data not used to modify it. It would record prior ChatGPT use, education, occupation, media exposure and access, then separate attitudes towards specific tasks from general beliefs about AI. Repeated surveys could show whether literacy precedes attitude change; randomized educational interventions could test whether teaching particular concepts changes calibrated trust without simply increasing approval.

Behavioural outcomes matter too. Researchers could compare stated concern with verification behaviour, willingness to delegate consequential tasks, ability to identify errors and actual learning performance. That would reveal whether the four attitude dimensions predict safe or unsafe use. Until such evidence arrives, this paper is best treated as a useful map of questions and associations. It shows that convenience samples in two European countries described ChatGPT differently and that literacy and broader beliefs travelled with those differences; it does not explain national cultures or prove how to change public trust.[1]