Students who felt that generative AI made them more capable also tended to report greater trust in the technology and greater AI dependence, according to a new peer-reviewed study of university students in China. The Frontiers in Psychology study surveyed 360 students with previous experience using generative AI.
The result is an association, not proof that feeling empowered by AI causes dependence. The study was cross-sectional and based on self-reported attitudes and behaviour. That limitation is central because trust, perceived usefulness and reliance can influence one another in several directions.
Feeling more capable and relying more were linked
The researchers measured four ideas: perceived AI empowerment, perceived AI threat, trust in AI and dependence on generative AI. Empowerment meant the sense that AI helped a student perform tasks more effectively, while dependence captured reported reliance on the technology rather than simply frequency of use.
Perceived empowerment was positively correlated with trust and dependence. Students who saw AI as more threatening tended to report less trust and less dependence. Trust itself showed a positive relationship with dependence and statistically accounted for part of the relationship between the students’ appraisals and their reported reliance.
The pattern is intuitively plausible. A tool that repeatedly feels useful can become a default part of a workflow. Once a student expects it to save time, explain difficult material or help generate ideas, not using it can begin to feel like giving up an advantage.
That does not make dependence inherently harmful. Calculators, search engines and spelling tools are all dependencies in a broad sense. The more useful question is what capability is being delegated, whether the student can still perform or evaluate the task independently and whether reliance improves learning or merely removes friction.
The sample makes the result interesting but narrow
Data were collected between 24 April and 25 May 2026 from universities in several Chinese provinces and cities, including Henan and Hubei. The researchers initially collected 739 questionnaires and retained 360 after quality screening. Just over half of the final sample were women, and most participants were under 22.
AI use varied. The paper reports 20.3% as rare users, defined as no more than once a month; 41.7% used generative AI one to three times a week; and 38.1% reported using it at least four times a week.
Those details matter because the sample is not a general measure of students everywhere. It was a convenience-based survey of students who had already used generative AI. Cultural context, university rules, subject mix and the particular tools students use could all affect attitudes.
The authors explicitly say the findings should be interpreted as associations rather than causal effects. A student who already relies heavily on AI may come to describe it as empowering because it is embedded in daily work. Alternatively, students who first experience strong benefits may become more reliant. A one-time survey cannot establish the direction.
Trust appears to be part of the mechanism
One of the stronger statistical relationships in the study was between trust in AI and reported dependence. The researchers found a correlation of 0.602 between the two measures. In their model, trust accounted for part of the association between perceived empowerment and dependence, and part of the negative association between perceived threat and dependence.
That does not mean trust explains everything. The measures were collected from the same respondents at the same time, and the model depends on how those concepts were defined and measured. But it gives educators a useful question to ask: what makes a student decide that an AI answer is reliable enough to lean on?
LiveAIWire has previously examined how AI assistance can interact with human skill development. The new study adds another layer by suggesting that students’ subjective experience of becoming more capable may coexist with a stronger habit of turning back to the tool.
That is not necessarily a contradiction. A student may genuinely complete a task better with AI and still learn less about how to complete it alone. Or the opposite may happen: AI could provide explanations and feedback that help the student develop a stronger independent understanding. The outcome depends on how the tool is used.
Education policy is moving towards competence, not simple prohibition
The study’s practical implication is not that students should distrust useful technology. It is that effective AI literacy needs to include judgement about when to delegate and when to verify.
UNESCO’s AI competency framework for students takes a similar broad approach. It emphasises a human-centred mindset, ethics, understanding AI techniques and the ability to evaluate AI critically rather than treating use itself as the goal.
That kind of competence becomes more important as generative AI moves from an occasional shortcut to an always-available layer around study. A student who knows how to prompt but cannot judge an answer has learned operation without oversight.
LiveAIWire has also covered work on AI acting as an instructor. The educational question is increasingly less about whether AI is present and more about the role it occupies: tutor, editor, search tool, co-writer or substitute thinker.
AI dependence should be measured by what disappears when the tool is removed
Frequency is an easy metric, but the paper itself notes that frequent use is not the same as dependence. Someone can use AI several times a day while remaining capable of checking and replacing it. Another person may use it less often but be unable to complete a key task without it.
A more practical way to think about reliance is to ask what happens when the AI is unavailable. Can the student still plan the work, evaluate sources, identify errors and explain the reasoning? If the answer is yes, the tool may be extending capability. If those functions have quietly migrated into the model, the apparent productivity gain may hide a loss of resilience.
Social context matters too. LiveAIWire’s coverage of AI and changing social norms points to a wider issue: once classmates routinely use a tool, opting out can feel like self-imposed disadvantage even when a student has concerns.
The new study does not settle whether generative AI makes students stronger or more dependent. Its value is in showing that those two experiences may not be opposites. People can feel more capable with a tool precisely at the point when the tool is becoming harder to do without.
Educators have to separate useful support from invisible substitution
The study points towards a practical problem for universities. The same interaction can look productive from the outside while doing very different things for learning. Asking an AI to explain a difficult idea in another way may help a student build understanding. Asking it to produce the reasoning that the student was supposed to practise may remove the very effort the task was designed to create.
That means simple rules based on whether AI was used can miss the important distinction. The educational question is what role the tool played. Did it give feedback, supply examples and expose weaknesses in an answer, or did it become the source of the answer itself? Two students can both report using generative AI several times a week while developing very different levels of independent skill.
UNESCO’s student AI competency framework emphasises human-centred judgement and the ability to evaluate AI rather than treating technical use alone as competence. That direction fits the tension in the Frontiers study. A student can feel empowered because a tool makes difficult work easier while still needing the judgement to notice when convenience has become dependence.
Longer studies would show whether dependence changes with experience
The biggest unanswered question is direction. A cross-sectional survey provides a snapshot of people who already differ in their habits, confidence and attitudes. Following the same students over time would make it easier to see whether an early positive experience increases later reliance, whether heavy reliance increases trust, or whether both are driven by a third factor such as workload or confidence with the subject.
It would also help distinguish temporary dependence from a durable change in study behaviour. Students often adopt new tools intensely during a demanding period and then change their habits once the novelty fades or course requirements shift.
For now, the study is most useful as a warning against treating empowerment and dependence as opposites. A technology can genuinely make somebody more capable in the moment and simultaneously become harder to work without. Education policy has to make room for both possibilities, encouraging tools that improve understanding while preserving enough independent practice for students to know when the AI is wrong.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.