On October 22, 2026, Oregon State University researchers reported that STEM students who trust and routinely use generative AI show significantly lower cognitive engagement — feeding a self-reinforcing "cognitive debt cycle." Here's what the study found, why our brains default to the easy path, and how to use AI without hollowing out your thinking.
Key takeaways
- The finding: In a survey of 299 STEM students at five North American universities, routine generative AI use was linked to weakened intellectual habits — and most surprisingly, the most "AI-savvy" students were spiraling deepest into dependence. (Oregon State University newsroom, Oct 22, 2026)
- The mechanism: Human thinking follows a "law of least effort" — deliberate, effortful thinking (system 2) yields to automatic shortcuts (system 1) whenever it can. Generative AI supercharges that shortcut because it delivers fluent, confident answers on demand.
- The loop: Weaker thinking habits lead to more AI reliance, which weakens thinking further — what the researchers call a self-reinforcing cognitive debt cycle.
- The fix: Use AI as a "bicycle for the mind" (amplifying cognition while you steer), not a substitute. A staged approach — think first, use AI second — is backed by independent learning-science guidance published the same day.
The study: 299 STEM students, one uncomfortable loop
The research, led by computer science graduate student Rudrajit "Rudy" Choudhuri with advisor Anita Sarma, Margaret Burnett (distinguished professor of computer science) and Christopher Sanchez (associate professor of psychological science), was published in the journal Information and Software Technology. The team analyzed survey data from 299 STEM students across five North American universities, with support from the USDA National Institute of Food and Agriculture and the National Science Foundation.
The core finding is a feedback loop. Routine use of generative AI serves to weaken students' intellectual habits, which leads them to lean on AI even more, further diminishing their thinking skills. As Choudhuri put it: "The most surprising finding was that the 'AI-savvy' ones, who you might think do significantly better in STEM careers, are actually spinning themselves deeper into the AI dependence spiral."
In the survey, students who trusted and routinely used generative AI reported significantly lower cognitive engagement. Sanchez summarized the behavior bluntly: students "are not using AI as an assistant but as more of a substitute for actually engaging with learning a topic." The researchers warn that "absent intervention, STEM education runs the risk of cultivating a generation of students unwilling and/or unable to exercise their capacity for reflection, understanding and critical thinking."
Why the brain takes the easy way out
The researchers anchor the mechanism in well-established cognitive psychology. Human ancestors who conserved mental energy survived; those who exhaustively analyzed every decision didn't. We learned to rely on intuitive, automatic, low-effort "system 1" thinking — the system that instantly answers 2 + 2 or completes the phrase "bread and ___" — and to reserve deliberate, controlled, effortful "system 2" thinking (calculating 17 × 24, evaluating a complex argument) for when it's unavoidable. "System 2 is lazy," Choudhuri said. "It avoids effort whenever System 1 can provide a plausible solution. It's what psychologists call the law of least effort, which is a pervasive tendency to accept the path of least cognitive resistance."
We've always offloaded thinking: shopping lists instead of memory, calculators instead of arithmetic, GPS instead of map reading. As Sarma noted, generative AI amplifies this trend "because they deliver fluent, confident support on demand. That means interacting with AI is more about choosing from outputs rather than thinking by doing."
That distinction matters for learning. STEM education depends on what psychologists call "desirable difficulties" — slow, painstaking cognitive work that builds intuition and transferable skills. Choosing from AI outputs feels productive through time savings, but it skips the struggle that makes knowledge stick.
Think first, AI second: the staged-use playbook
Independent guidance published the same day points in the same direction. Writing in The Conversation on October 22, 2026, learning-science researcher Luke Rowe — author of a 2026 paper in Education Sciences on the "Gradual Release of Technology" (DOI: 10.3390/educsci16081291) — argues that students should master foundational knowledge and skills before using AI to assist their learning.
His practical protocol:
- Before asking AI anything, retrieve, explain, solve, summarize or write the concept yourself first.
- Then use AI to organize your notes, spot gaps, and fact-check — or to generate progressively harder practice questions you attempt unaided.
- Use AI for feedback, not answers: "Don't rewrite it for me. Identify what I did well, what is inaccurate or missing, and the three changes that would most improve my answer."
- Never practice exams with AI assistance — Rowe notes that AI can temporarily boost performance, producing a false sense of confidence, followed by what he calls "performance collapse" on the real exam. He cites a June 2026 study of high school students that found AI use improved homework scores while monthly exam scores dropped about 20% within six months.
- His rule of thumb: "think first, use AI second (or, in some cases, not at all)."
AI as assistant vs. AI as substitute
| Study task | AI as substitute (erodes learning) | AI as assistant Better for learning |
|---|---|---|
| Practice exams | Chat with AI open beside the paper | Closed laptop, timer on, retrieve unaided |
| First draft of an essay | AI writes the draft from a prompt | You draft; AI critiques against the rubric |
| Learning a new concept | Ask AI to explain; accept the answer | Explain it yourself first; use AI to find gaps |
| Debugging code | Paste the error, apply the fix | Attempt the fix, then ask AI to review your reasoning |
| Reviewing notes | Ask AI for a summary | Summarize yourself; ask AI to quiz you on it |
The Oregon State team's design suggestion matches this table's spirit: build AI tools as "bicycles for the mind." A bicycle amplifies physical efficiency while requiring the rider's active control; a mind-bicycle would amplify cognition via critique, counterargument and provocation while keeping humans responsible for exploration, steering and judgment.
The verdict: which learner are you?
- Student using AI daily for coursework → adopt the think-first protocol — attempt retrieval, drafting and problem-solving unaided before opening the chatbot; use AI for feedback and progressively harder practice questions.
- Teacher or lecturer who can't ban AI → redesign the assessment, not the rules — alternate AI-assisted and independent work in supervised settings, and make in-class assessment and collaborative tasks carry real weight so effort pays off.
- Professional self-learner upskilling with AI → treat fluency as a warning sign — if the session felt effortless, you probably farmed out the desirable difficulties; redo the hard parts without assistance.
The studies don't argue that AI is the enemy of learning. They argue that effort is the mechanism of learning — and that the most important AI literacy may be knowing when to close the laptop.
Sources
- Oregon State University Newsroom — "Tech-savvy STEM students prone to letting AI do their thinking for them, study shows," October 22, 2026 (study details, researcher quotes; study led by Rudrajit "Rudy" Choudhuri, published in Information and Software Technology; survey of 299 STEM students at five North American universities)
- The Conversation / phys.org — Luke Rowe, "Are you using AI to study? Here's how to do it well," October 22, 2026 (Gradual Release of Technology approach; Education Sciences (2026), DOI: 10.3390/educsci16081291)
- Featured photo: Unsplash — study desk with notebook, laptop and coffee (https://images.unsplash.com/photo-1499750310107-5fef28a66643)
Comments (0)