When OpenAI released "study mode" in July 2025, the company's education vice president told reporters the tool "can significantly improve academic performance." The claim lands in a classroom reality where AI adoption is racing ahead of the research meant to evaluate it. A handful of studies show benefits for specific groups like programming students and English language learners, but other research points to impaired performance and weakened critical thinking when the tools are taken away.
One paper found that the more a student used ChatGPT while learning, the worse they performed later on similar tasks without it. Scholars have also flagged methodological weaknesses in many optimistic studies, including one published in Nature in May 2025 that suggested chatbots may aid higher-order thinking. The evidence base is thin, and the long-term picture remains unclear.
Two systems of thinking explain why shortcuts backfire
Cognitive psychology distinguishes between two processing modes. System 1 is fast, automatic, and driven by pattern matching and habit - getting dressed or riding a familiar bike route. System 2 is slow, deliberate, and requires conscious mental effort. Gaining knowledge and mastering new skills depends heavily on System 2.
"Struggle, friction and mental effort are crucial to the cognitive work of learning, remembering and strengthening connections in the brain," the research shows. A confident cyclist relies on System 1 pattern recognition built through hours of effortful System 2 work. Without that initial strain, mastery does not develop.
Offloading mental work creates metacognitive errors
When a machine does the cognitive heavy lifting, the brain loses its workout. Research on GPS use has demonstrated that habitually offloading navigation can impair spatial memory. Using Google to answer questions makes people overconfident in their own knowledge - a phenomenon psychologists call metacognitive error.
One study found that students researching a topic with ChatGPT experienced lower cognitive load during the task but produced worse reasoning afterward. Another showed that students who used AI to revise essays scored higher by copying and pasting sentences, yet showed no actual knowledge gain compared to peers who worked without it. The AI group engaged in fewer rigorous System 2 thinking processes, and the study's authors warned that "metacognitive laziness" may produce short-term performance bumps at the cost of long-term skill stagnation.
For educators looking to understand these dynamics more deeply, the AI Learning Path for Teachers examines how cognitive load theory applies to classroom AI use.
AI tutors show promise - and introduce new problems
AI companies are designing tools meant to function as Socratic tutors. Anthropic released its learning mode for Claude in April 2025, and OpenAI's study mode is positioned similarly - posing questions and providing hints rather than answers. But early results are mixed.
In one study, high school students reviewing math with ChatGPT performed worse than those who studied without AI. Some used a customized tutor version designed to give hints without revealing answers. When tested later without AI access, those who had used the base version of ChatGPT did much worse than the no-AI group - and did not realize their performance had dropped. Students who used the tutor bot performed no better than those who reviewed without AI, but mistakenly believed they had done better. The AI did not help, and it introduced the same metacognitive errors seen in other offloading research.
Even as tutor modes improve, students must actively select that mode and guide the chatbot away from low-level questions or sycophancy. Those design problems may be fixable. The deeper challenge is the temptation to use default-mode AI to bypass hard work - a classic problem of course design and student motivation.
Why this matters for educators
The research does not say AI has no place in learning. Offloading can be useful once foundations are in place, but those foundations cannot form unless the brain does the initial encoding and connection work. The gym metaphor is direct: a personal trainer who pushes you to work harder is valuable; a robot that lifts the weights for you causes atrophy. Educators deciding when and how to introduce AI tools need to distinguish between the two. For training resources that address this distinction directly, AI for Education Training covers practical approaches to integrating AI without undermining the cognitive effort that produces real learning.
Your membership also unlocks:
