In one of the largest studies on AI and cognition to date, the correlation between frequent AI use and critical thinking scores was r = -0.68. For context, that's a stronger relationship than the one between smoking and lung cancer in many early epidemiological studies. The students who used AI the most scored the lowest on critical thinking, by a wide margin.
We don't think AI is poison for young minds. But we do think these numbers should change how schools approach AI integration, and quickly.
The Gerlich Study
The most widely cited study comes from Michael Gerlich at SBS Swiss Business School, published in January 2025 in the journal Societies. Gerlich surveyed 666 participants across three age groups in the United Kingdom and found a strong negative correlation (r = -0.68) between frequent AI tool usage and critical thinking abilities.
The mechanism? Cognitive offloading, the tendency to let external tools do the mental work that would otherwise strengthen cognitive skills. When a student asks ChatGPT to summarize an article instead of reading and synthesizing it themselves, they skip the cognitive effort that builds comprehension and analytical ability.
Key findings from the study:
- Younger participants (17-25) exhibited higher dependence on AI tools and lower critical thinking scores compared to older groups
- Cognitive offloading mediated the relationship: the more participants outsourced thinking to AI, the lower their critical thinking scores
- The effect was consistent across educational backgrounds
An important caveat: the study shows correlation, not causation. It's possible that people with lower critical thinking skills are more drawn to AI tools, rather than AI tools causing the decline. But the strength of the correlation, combined with the cognitive offloading mechanism, suggests the relationship is at least partially causal.
The Broader Evidence
Gerlich's study isn't an outlier. Multiple lines of research converge on the same conclusion:
Adoption Is No Longer the Question
According to Pew Research Center data, the share of U.S. teens using ChatGPT for schoolwork doubled from 13% to 26% between 2023 and 2024. Among college students, Campus Technology reports that 86% use AI in their studies. A College Board survey found that 84% of high school students used generative AI tools for schoolwork by mid-2025.
The adoption question is settled. What matters now is whether students are using AI in ways that support or undermine learning.
Why Struggle Matters: The "Generation Effect"
Decades of learning science research explain why cognitive effort matters. The generation effect, first documented by Slamecka and Graf (1978), shows that people remember information better when they generate it themselves rather than passively receiving it. When a student writes their own summary, they engage in retrieval practice, elaborative encoding, and organizational processing, all of which strengthen memory and understanding.
Related research on desirable difficulties (Bjork & Bjork, 2011) shows that learning conditions that feel harder often produce better long-term outcomes. Struggle isn't a bug in the learning process. It's a feature. When AI removes that struggle, it may also remove the learning.
Writing as Thinking
The connection between writing and thinking is well-established in educational research. Writing is not simply recording pre-formed thoughts. It's a cognitive process that generates thinking. As the National Council of Teachers of English (NCTE) has long maintained, writing is a tool for discovery, analysis, and meaning-making.
When students outsource writing to AI, they don't just skip the writing. They skip the thinking that writing produces. A student who asks ChatGPT to write their analysis of The Great Gatsby doesn't learn less about essay structure; they learn less about The Great Gatsby.
What Teachers Can Actually Do
None of this means banning AI. It means being deliberate about when and how students use it:
1. Make the Process Visible
Instead of evaluating only final products, track and assess the writing process. When students know that their revision history, drafts, and writing patterns are part of the evaluation, they're incentivized to do the cognitive work themselves.
Process-based assessment tools can show teachers:
- How a student's essay evolved through multiple drafts
- Where pauses, revisions, and restructuring occurred
- Whether text was typed keystroke by keystroke or pasted in large blocks
- The total time and effort invested in composition
2. Design AI-Resistant Assignments
Some assignment types are harder for AI to complete meaningfully:
- Personal reflection connected to specific classroom discussions or experiences
- Iterative revision where students submit multiple drafts with reflection on changes
- Source-specific analysis requiring engagement with particular texts assigned in class
- In-class writing components that establish a baseline for each student's voice
3. Use AI as a Learning Partner, Not a Shortcut
The goal isn't to eliminate AI but to use it in ways that enhance rather than replace cognitive effort:
- Students can use AI to generate counterarguments to their own thesis, then respond to them
- AI can serve as a Socratic tutor, asking questions rather than providing answers
- Students can evaluate and critique AI-generated responses, building critical analysis skills
4. Teach Metacognitive Awareness
Help students understand why doing the work matters. When students grasp the connection between effort and learning, they're more likely to use AI as a supplement rather than a substitute. The APA Style Guide's recommendations on teaching writing in the AI era emphasize making students conscious of their own cognitive processes.
The Adaptation Gap
A Frontiers in Education synthesis of research from 2023-2025 found that while generative AI is transforming academic reading and writing, most teachers have not yet significantly adjusted their assessment practices. Students are sprinting ahead with AI adoption; assessment is standing still.
AI tools aren't inherently harmful to learning, but passive, unstructured use appears to be. The research keeps landing in the same place: the cognitive effort of writing, thinking, and struggling with ideas is where learning happens. When AI shortcuts that effort, it shortcuts the learning itself.
We think the fix is straightforward, even if it's not easy: design educational experiences where the process matters as much as the product, and use tools that make that process visible and verifiable. That's what Limudi does. It shows teachers the complete writing process, every keystroke, revision, and pause, so authentic student effort is never in question.