Why students use ChatGPT—and what trust conceals

The study examined actual usage behavior

Many acceptance studies measure only whether people intend to use a technology in the future. Polyportis and Pahos extended a meta-UTAUT framework and included actual usage behavior as an outcome.

Using structural equation models, relationships between attitude, intention, usage, and additional factors were tested. The data came from 355 students in the Netherlands.

Attitude and behavioral intention were significant positive predictors of usage. This confirms that a positive attitude is important but does not suffice on its own.

The study is observational. Statistical associations show a plausible model but do not prove a clear causal role for every factor.

Benefits, context, and opportunities shape intent

Performance expectancy, social influence, and facilitating conditions had a positive effect on behavioral intention. Students are more likely to use a system when they expect a concrete benefit, when others normalize it, and when access is practically feasible.

Expected effort also plays a role through attitude. An easily accessible chat lowers the threshold for asking questions, checking texts, or gathering initial ideas.

However, these factors do not explain whether the answer is correct or whether use is permitted. Adoption and quality are separate levels.

For product teams, this is a warning against vanity metrics. Many sessions can arise from ease of use, novelty, or social pressure and are not yet evidence of a good substantive effect.

Anthropomorphism and novelty improve attitude

Perceived humanness, novel design, and trust were among the positive antecedents of attitude. A system appears more attractive when it seems socially familiar and at the same time technically new.

Anthropomorphism can emerge through language, name, avatar, and interaction rhythm. It may facilitate access, but it also creates expectations of understanding and reliability.

Design novelty is often limited in time. What initially seems surprising quickly becomes the normal interface. Long-term use must rely more on actual benefit and consistent behavior.

A responsible companion can be designed to be memorable without suggesting consciousness or a human relationship. The AI labeling should remain unambiguous despite social design.

Institutional rules change the step toward use

Institutional policy negatively moderated the relationship between behavioral intention and actual use. Even with positive intention, rules can limit practical use.

In the higher education context, this concerns, for example, permitted aids, exam integrity, data protection, and labeling of AI support. Clear rules provide orientation but can also restrict use to specific tasks.

For mental or health-related chats, organizational rules correspond to defined roles and care pathways. Not every technically possible response should become visible or action-guiding in the product.

Good governance does not merely explain prohibitions. It describes what the system is intended for, how uncertainty is handled, and who takes responsibility for errors.

Trust leads to intention and use

Choudhury and Shamszare surveyed 607 U.S. adults who used ChatGPT 3.5 at least monthly. Trust had significant direct effects on intention to use and actual use.

The model explained 50.5 percent of the variance in intention to use, but only 9.8 percent of actual use. Thus, many further practical and situational influences lie between trust and behavior.

The authors emphasize the risk of overtrust in health questions, because ChatGPT was not originally developed as a health system. At the same time, too little trust can prevent useful applications.

The target value is therefore not maximum trust. It is an appropriate assessment of what the specific system can achieve in the specific task.

Banking chatbots reveal context-specific trust factors

Alagarsamy and Mehrolia evaluated 435 complete questionnaires from users of four major Indian banking chatbots. The tested antecedents explained 38.6 percent of the variance in trust.

Interface, design, and technophobia were not among the significant explanatory factors in this model. Trust again explained smaller portions of attitude, intention, and satisfaction.

The pattern differs from the higher education study, in which design novelty and anthropomorphism were related to attitude. Thus, target group, task, and examined constructs influence which factors become visible.

There is no universal recipe for chatbot trust. Research findings should be read in their respective contexts rather than being directly transferred from e-commerce, banking, or education to mental health.

Adoption is not evidence of quality

A product may be used frequently because it is free, fast, and socially visible. These attributes are valuable but say little about factual accuracy, conversational boundaries, or long-term impact.

For supportive chats, two separate measurement systems are therefore needed. Usage data show whether people find access and return. Quality tests show whether the system maintains the chosen direction, incorporates user corrections, and does not reinforce inappropriate interpretations.

User feedback should additionally capture why someone returns. A high number of messages may indicate satisfaction, but it could also reflect a tedious search for a suitable answer.

Discontinuation is also ambiguous. Perhaps the person was finished, perhaps disappointed, or technically blocked. Numbers require cautious interpretation and qualitative supplementation.

For young target groups, it should also be examined whether usage differences are related to age, education, language, or digital experience. A high overall average can obscure groups that barely reach the offering or trust it too much.

Responsible design combines attractiveness and boundaries

The university study shows how attitude, trust, social influences, design, and institutional rules together shape usage. It explains why a good service is more than its model.

For a conversational partner, a low entry barrier, a natural tone, and a recognizable space make sense. At the same time, the role, data path, and AI character must be clear.

Good governance does not stand as a legal block alongside the product. It becomes visible in behavior: the chat accepts a no, separates opinion from facts, and only refers onward when the task requires it.

Why people use ChatGPT does not automatically explain whether the usage helps them. Only when adoption, quality, and consequences are examined together does a popular chat become a responsibly developed system.