Studies find college students lean on ChatGPT but rarely check its answers

A study from the University of Granada examined how final-year undergraduates in Education and Information Science use ChatGPT, drawing on a sample of 50 students. Researchers combined a focus group using think-aloud methods, semi-structured interviews, and a Likert-scale questionnaire, then compared the quantitative and qualitative results against each other to check consistency. The goal was to understand how ChatGPT use relates to information literacy, the set of skills involved in finding, evaluating, and using information effectively.

A separate study surveyed 880 university students in southwestern China, spanning undergraduate, master's, and doctoral levels, with an average age of 20.8 and a gender split of 48.52% male. It combined structural equation modeling, a statistical method for testing relationships between survey-based factors, with an explainable machine learning component, an approach designed to show which factors most influenced a model's predictions. Cornell University's Center for Teaching Innovation, meanwhile, has been synthesizing emerging research on generative AI's classroom effects to guide its own faculty, rather than running a study of its own.

At Granada, students reported using ChatGPT mainly for retrieving information, but they often did so without clearly defining their questions or what they actually needed to know. They found it useful for processing text: summarizing passages, pulling out keywords, sorting material into categories, and drawing connections between ideas. Where they struggled was references: students described the tool's handling of citations as confusing and inaccurate, and they said they wanted more training in how to use it well, along with faculty who were themselves more familiar with the technology.

The China-based survey found a similar mix of pull and hesitation, framed in terms of what draws students toward AI tools and what holds them back. Perceived usefulness and personal interest strongly encouraged adoption, while perceived risks tied to ethics, accuracy, and academic integrity significantly discouraged it. The researchers describe this as a benefit-risk balance: students weigh what the tool can do for them against worries about whether its output can be trusted or whether using it crosses a line into dishonesty.

Cornell's synthesis of the wider research points to a related pattern in how ChatGPT and similar large language models actually work. These systems, called large language models, are trained on huge amounts of text and generate responses by predicting the most likely next words, rather than by understanding facts. Because of this, they can state false information as if it were true, a failure researchers call a hallucination. Cornell's materials note that while ChatGPT can answer questions and explain concepts, it does not have knowledge or understanding of what it generates in the way a person does.

Cornell's review of outside research on generative AI's classroom effects describes genuinely mixed results. Some studies found learning gains from student use of AI tools, particularly in math or computer science courses, but those gains often proved temporary rather than lasting, according to two cited studies. Other research on writing was described as more concerning: one study found that students using generative AI for writing tasks showed less brain activity, produced more homogeneous writing, and were less able to recall their own written work afterward compared to students who did not use it.

At the same time, Cornell notes that some research is beginning to focus on ways generative AI might support learning rather than undermine it, such as helping students practice actively, giving them actionable feedback, or prompting them to reflect on their own learning process. Some AI tools have introduced dedicated "study modes" built around these approaches. Cornell frames this as an open and ongoing question, saying it will continue to let new research inform how it advises faculty, rather than treating the matter as settled.

None of the sources claim to have settled how generative AI affects learning overall. The Granada researchers conclude that students want more structured training in AI tools and more faculty familiarity with the technology, framing current use as underdeveloped rather than harmful outright. The China-based team frames its findings as a snapshot of what drives adoption and hesitation among the students it surveyed, not a verdict on outcomes.

Cornell's own framing is explicitly provisional: it states that research into generative AI's impact on learning is still emerging, with new results appearing every month, and that some of the studies it draws on are themselves recent preprints or working papers rather than fully settled findings. A fourth line of inquiry, on whether large language models process language consciously or unconsciously, raises separate questions about what these systems are actually doing computationally when they produce text that resembles human writing, though this work does not address classroom use directly.