As generative AI becomes more prevalent in education, many see it as a promising assistant for students navigating complex academic tasks. But a new study titled „Generative AI Can Harm Learning: Five Experiments“, authored by Chinmayi Sharma, Anne Spencer Ross, and Paul V. Timmers and published in Science, raises serious concerns about its unintended consequences. The article presents a series of carefully designed experiments showing that while generative AI tools like ChatGPT can produce helpful-looking answers, they often impair learning outcomes and reduce students‘ engagement with the material.
Across five experiments involving over 4,000 university students, the researchers systematically examined how AI assistance influenced performance on reasoning tasks, motivation to study, and retention of knowledge. Participants were asked to complete assignments ranging from writing tasks to logic problems, with some having access to generative AI and others working independently. The results were surprisingly consistent: students who relied on AI performed worse on subsequent assessments and demonstrated weaker grasp of the core concepts. This pattern held true even when the AI’s responses were accurate or superficially helpful.
The authors attribute this phenomenon to a key psychological mechanism: cognitive offloading. When students use AI to handle thinking-intensive tasks, they skip the deep mental processing that is essential for learning. By outsourcing the cognitive work, they weaken their ability to internalize the information. In some cases, students even developed overconfidence, assuming that because their answers looked polished, their understanding must be solid—an illusion that became apparent during later testing.
Another important finding is the role of perceived effort. Students who used generative AI reported lower motivation to study further and engaged in fewer self-correction behaviors. Even when prompted to check the AI’s output, many participants failed to spot subtle errors or logical flaws, indicating a dangerous level of trust in machine-generated content. This suggests that the availability of generative AI may not only reduce learning but also diminish critical thinking habits over time.
Interestingly, the study also explored whether AI could be harnessed in a more productive way. In one of the experiments, students were asked to critique and correct the AI’s answers rather than use them passively. This interactive approach showed modest improvement in learning outcomes, suggesting that if generative AI is framed as a tool for engagement rather than substitution, it could have educational value. However, the authors caution that this requires deliberate instructional design and cannot be left to student discretion alone.
The researchers make it clear that their results should not be seen as a rejection of generative AI in education. Rather, they call for a more nuanced integration that emphasizes human cognition and avoids overdependence. Educators must rethink assessment design, embed metacognitive prompts, and promote transparency about how AI tools are used. Without these adjustments, the risks may outweigh the benefits, particularly in subjects that rely heavily on problem-solving and conceptual mastery.
The implications of this research extend beyond classrooms. As generative AI becomes a routine feature in workplaces and daily life, understanding its cognitive effects will be essential. If people rely on it for every complex decision, the erosion of independent thinking could have long-term consequences for professional competence, democratic participation, and innovation.
The full article by Chinmayi Sharma, Anne Spencer Ross, and Paul V. Timmers, titled “Generative AI Can Harm Learning: Five Experiments” is available online here.