Generative artificial intelligence is rapidly becoming part of higher education, with tools such as ChatGPT, Google Gemini, Claude, and Microsoft Copilot now widely used by university students for studying, writing, reviewing concepts, and solving academic problems.
Across Latin America, the adoption of generative AI among students has grown quickly. In 2026, about 92 percent of students surveyed by the Digital Education Council reported using generative AI tools. This rapid shift has created an urgent challenge for universities: how should they respond to AI use in learning environments?
Many institutions are still deciding whether to restrict, regulate, or integrate AI tools into teaching and assessment. While student use of AI continues to expand, rigorous evidence on its actual impact on learning is still developing. This creates a gap between practice and policy, as universities must make decisions before the full academic effects of AI are clearly understood.
A recent experimental study in Chile provides important evidence on this issue. The study examined whether universities can influence how students use AI tools and whether that guidance can affect learning outcomes. The research focused on undergraduate econometrics students at a selective university in Chile and used a course-specific AI assistant called GPT-U.AI.
The study tested two randomized interventions. The first intervention encouraged students to adopt the AI tool before a midterm exam. Students in the treatment group received three emails introducing GPT-U.AI, explaining how it could be used, and sharing a short instructional video. They were encouraged to ask questions about concepts and use the tool for exam review.
Students in the comparison group also had access to GPT-U.AI, but they did not receive the instructional messages. This allowed the researcher to measure whether encouragement alone could increase adoption and improve academic performance.
The results showed that encouragement increased awareness and use of the AI assistant. Students who received the emails were more likely to know about GPT-U.AI and more likely to report using it. Their reported usage intensity also increased.
However, greater adoption did not lead to better midterm exam performance. The intervention increased exposure to the tool, but it did not significantly improve learning outcomes. One possible reason is that students did not view GPT-U.AI as especially valuable compared with the many other study resources already available to them.
The study found no meaningful difference in students’ perceptions of the tool’s value after the first intervention. This suggests that simply introducing students to an AI tool may not be enough to change how they learn. In a context where students already have broad access to generative AI, awareness alone may have limited impact.
The second intervention shifted the focus from adoption to quality of use. Before the final exam, students in the treatment group received three additional emails encouraging them to use GPT-U.AI in a more learning-oriented way. Instead of simply recommending the tool, the messages guided students to use the assistant as a tutor.
The guidance encouraged students to ask the AI assistant to explain econometrics concepts step by step, support reasoning rather than provide final answers immediately, generate practice questions, provide feedback, verify assumptions, and help identify mistakes in their thinking.
This second intervention produced stronger results. Students who received guidance on how to use AI improved their final exam performance by 0.22 standard deviations. The gains were concentrated in open-response sections that required structured reasoning and problem-solving, rather than in multiple-choice questions.
This distinction is important because the guidance did not focus on memorization or answer retrieval. Instead, it encouraged students to interact with AI in a way that supported reasoning, reflection, and deeper understanding. The findings suggest that AI can improve learning when students are guided to use it as a tutor rather than as a shortcut.
The study also showed changes in how students interacted with the AI assistant. Students who received the guidance were more likely to use GPT-U.AI in tutor mode, engaging in step-by-step learning rather than simply asking for direct answers. The intervention also increased students’ perceived usefulness of the assistant by 0.38 standard deviations.
Importantly, the guidance did not increase overall use of generic AI tools such as ChatGPT, Gemini, or Claude. These tools were already widely used by students. Instead, the intervention changed how students used a course-aligned AI assistant, showing that institutional guidance can shape the quality of AI use even when access to AI is already widespread.
The findings have important implications for universities. Many institutions have focused heavily on preventing or restricting AI use, especially because of concerns about plagiarism, academic integrity, and overreliance on automated tools. While these concerns are valid, the evidence from Chile suggests that universities may also have an opportunity to guide students toward more productive uses of AI.
AI does not automatically improve learning, and access alone may not be enough. However, when students are encouraged to use AI as a tutor that supports reasoning, feedback, and step-by-step problem-solving, the technology may complement cognitive effort rather than replace it.
This is especially relevant because the interventions used in the Chile study were relatively low-cost. The university did not need major infrastructure, expensive licenses, or a full redesign of the course. The guidance was delivered through structured prompts using the existing course platform and standard email communication.
For universities operating with limited budgets, this is an important lesson. Improving the educational value of AI may not always require large investments. In some cases, carefully designed guidance, course-specific prompts, and clear expectations for learning-oriented use may be enough to make AI more beneficial.
The study also contributes to the growing body of research on generative AI in education. It shows that the impact of AI depends not only on whether students use it, but also on how they use it. Tools that can either support learning or weaken effort must be shaped by thoughtful institutional guidance.
The results should not be assumed to apply automatically to every context. The study focused on undergraduate econometrics students at one selective university in Chile. More evidence is needed across different disciplines, universities, student populations, assessment types, and teaching models.
Even with these limitations, the findings point to a clear lesson for higher education. Universities should not treat AI use as only a problem to control or a technology to provide. They should treat it as a learning practice that requires guidance, structure, and evaluation.
As large language models become more embedded in student life, the role of universities will be critical. Institutions can help students understand when AI supports learning, when it may undermine independent thinking, and how to use it responsibly in academic work.
The evidence from Chile suggests that guiding college students’ AI use can improve learning, especially when the guidance encourages deeper reasoning and active engagement. Access to AI may be widespread, but educational value depends on whether students are taught how to use these tools well.
For universities around the world, the message is simple: AI adoption is not only about availability. It is about direction. When institutions guide students toward thoughtful, tutor-like interaction with AI, generative tools may become a meaningful support for learning rather than just another digital shortcut.
