As large language models (LLMs) like ChatGPT become more embedded in academic, professional, and creative workflows, the question of user trust is not just timely—it’s essential. Understanding what drives or erodes trust in AI tools can guide ethical deployment, user education, interface design, and societal adaptation.
This article unpacks the findings of a mixed-methods study conducted among UK university students, exploring four major domains shaping trust in ChatGPT: (1) user attributes, (2) seven core trust dimensions, (3) task context, and (4) societal perceptions of AI. By combining survey responses from 115 students with insights from semi-structured interviews, the study offers nuanced perspectives on how trust is formed, challenged, and reshaped in real-world educational contexts.
One of the most surprising findings was that demographics like age, gender, and academic discipline had minimal impact on trust levels. Instead, behavioral engagement was the primary predictor:
Frequent use of ChatGPT correlated with higher trust.
In contrast, a deeper self-reported understanding of LLM mechanics (how ChatGPT works under the hood) actually decreased trust.
This paradox suggests a phenomenon observed in other tech trust domains: the more you know, the more critical you become. Students who engage with the limitations of LLMs—such as hallucination, bias, or lack of reasoning—tend to approach outputs with healthy skepticism. Meanwhile, regular users who rely on ChatGPT in a practical way may build experiential trust, focusing on the tool’s utility rather than its inner workings.
Notably, computer science students did not uniformly trust ChatGPT more than their peers. They only showed elevated trust when using it for proofreading and writing—contexts where domain expertise aligns with evaluating output quality.
🔍 Takeaway: Trust is cultivated more by what users do with ChatGPT than by who they are. Frequent, successful interactions build confidence, while technical awareness tempers it.
The study mapped seven dimensions of trust, adapted from prior research on human-computer interaction and AI ethics:
Perceived Expertise
Ethical Risk
Ease of Use
Transparency
Human-Likeness
Reputation
Reliability
Of these, perceived expertise (i.e., “Does ChatGPT know what it’s talking about?”) and ethical risk (“Could ChatGPT be misused or cause harm?”) were the strongest predictors of overall trust.
Ease of use and transparency had moderate but meaningful effects.
Surprisingly, human-likeness and reputation were not significant.
This challenges assumptions that making AI “more human” increases trust. On the contrary, students may separate interface design from functional credibility. Many appreciated ChatGPT’s fluency and politeness, but didn’t equate that with actual expertise or trustworthiness.
Likewise, ChatGPT’s brand reputation, while globally recognized, did not independently elevate trust. Students seem to evaluate trust based on personal experience and task performance, not name recognition.
🔍 Takeaway: Trust in AI is shaped by a complex interplay of perceptions. Competence and ethics matter more than polish or personality.
Perhaps the clearest pattern in the study was that trust is highly task-dependent. Students reported varying levels of trust in ChatGPT depending on the type of task they used it for:
| Task Type | Trust Level |
|---|---|
| Coding assistance | High |
| Summarization | High |
| Proofreading | Moderate |
| Creative writing | Moderate |
| Entertainment content | Low |
| Citation generation | Very Low |
Despite the low trust in citation generation, the most startling finding was that confidence in ChatGPT’s referencing ability—regardless of its actual accuracy—was the strongest correlate of overall trust. This is a classic example of automation bias, where users over-rely on technology even when they know it makes mistakes.
🧠 Automation bias occurs when users defer judgment to an automated system, assuming it's more accurate or capable than it really is.
This shows that even informed users may suspend disbelief when AI appears authoritative—particularly in academic referencing, where correct citations signal legitimacy. The danger here is obvious: hallucinated references could easily mislead students or compromise academic integrity.
🔍 Takeaway: Trust must be interpreted through the lens of use-case specificity. AI may be ideal for some tasks but dangerously flawed for others.
Beyond individual tasks or features, students’ broader beliefs about AI's societal impact significantly shaped their trust in ChatGPT:
Those with a positive view of AI’s role in education, productivity, and society were more trusting.
Those with a negative or ambivalent view expressed skepticism, even when they acknowledged ChatGPT’s usefulness.
This indicates that trust in ChatGPT is not just functional—it’s ideological. Students who fear job displacement, surveillance, or ethical erosion from AI systems may carry that skepticism into every interaction with LLMs. Meanwhile, techno-optimists are more inclined to see tools like ChatGPT as augmentative, not threatening.
Interestingly, students in creative fields often expressed a dual perspective: excitement about using AI to brainstorm or co-write, but concern that AI-generated content might dilute originality or lead to homogenized thinking.
🔍 Takeaway: Individual trust is embedded within broader societal narratives about AI—utopian or dystopian. These narratives color every interaction.
The study’s findings have actionable relevance across several domains:
Encourage critical use of ChatGPT, not blanket approval or prohibition.
Emphasize task-specific evaluations: e.g., it’s great for summarizing but not for sourcing citations.
Incorporate AI literacy modules to help students understand what ChatGPT can—and cannot—do.
Improve referencing capabilities, or issue clearer disclaimers when generating sources.
Prioritize transparency cues, such as evidence trails or confidence scores.
Avoid overhumanizing the interface; students aren’t swayed by chatbot “personality.”
Frame regulations that promote ethical safeguards without stifling innovation.
Address automation bias through interface design and public education.
Invest in research on AI’s educational role that accounts for diverse student populations and use cases.
This study underscores a fundamental truth: user trust in AI is multidimensional, contextual, and deeply human. It is shaped not only by the design and performance of ChatGPT, but also by who users are, what they’re trying to achieve, and what they believe about the role of AI in the world.
For university students, trust in ChatGPT emerges from a tension between utility and uncertainty, efficiency and ethics, fluency and fallibility. Rather than aiming to “maximize” trust in AI tools, we should instead foster calibrated trust—enough to use the tool effectively, but not blindly.
As we enter an era where LLMs are integrated into everything from lesson planning to legal advice, studies like this offer a roadmap for building systems—and societies—that are not only smart, but wise.
[Original Study Summary]
Parasuraman & Riley (1997), Humans and Automation: Use, Misuse, Disuse, Abuse
Hoff & Bashir (2015), Trust in Automation
Binns et al. (2018), ‘It's Reducing a Human Being to a Percentage’: Perceptions of Justice in Algorithmic Decisions
