For the lecturer who keeps reading that AI is eroding student thinking and wants to know what, specifically, to do about it in the next class.
Every fresh term, faculty face some version of the same question: are students learning, or just moving through the motions of a degree? This year that question has a new backdrop, and the numbers behind it are hard to wave away.
What faculty are afraid of
A November 2025 survey of 1,057 faculty by the American Association of Colleges and Universities and Elon University's Imagining the Digital Future Center found that 90% believe generative AI will diminish students' critical thinking skills. 83% believe it will shorten their attention spans, and 95% expect it to increase students' overreliance on the tools themselves. Nearly three in four, 73%, have personally dealt with an academic integrity issue involving AI.
These are not fringe worries from a resistant minority. Among the faculty who responded, this is the majority view, and it is specific. Not “AI is disruptive” in the abstract, but a shared sense that something measurable is happening to how students think.
The same tool, two different students
Students are not hiding from this either. HEPI's 2026 Student Generative AI Survey, based on 1,054 UK undergraduates, found that 95% now use AI in some form and 94% use it for assessed work. Only 12% admit to submitting AI-generated text directly, though that figure has quadrupled since 2024, when it stood at 3%. The line between “using AI to help” and “letting AI do it” is moving fast.
Two of the survey's own quotes, set side by side, tell the story better than any statistic. One student described using AI to summarize dense readings so they could “focus on critical analysis and deeper understanding.” Another wrote, simply, “I'm not using my brain at all.” Same tool, same term, two entirely different learning experiences. The survey gives no sign that either student was doing anything unusual.
Why the mechanism has not changed
What separates those two students is not the tool. It is what the classroom asked of them once they had used it.
AI can answer the “what” faster than any lecture, any textbook, any set of office hours. What it cannot do is stand in for a student at the moment they are asked to predict an outcome, defend a choice, or explain something back in their own words. Those moments only work if the student is the one doing the retrieving.
This is the older idea underneath the whole debate: retrieval practice, sometimes called the testing effect. It is the well-established finding that actively pulling information out of memory builds stronger, longer-lasting understanding than being handed the same information again. It is not new, and it is not about AI. It is just that AI makes it unusually easy to skip.
The 16-point difference
A 2025 study led by researchers at Drexel University (An et al.) puts a figure on the gap. In a data science course, students were given AI-generated retrieval-practice questions during some weeks, low-stakes prompts that forced them to recall and apply what they had just been taught, and none in others. The same students' quiz performance was then compared week to week.
They scored 89% in the weeks with retrieval practice and 73% in the weeks without. A 16-point difference the researchers called both statistically and practically significant, produced by changing nothing about the content, only whether students were made to retrieve it.
That finding does not change because AI exists. If anything, it names exactly what is at stake. The mechanism that produces real learning still runs on effortful recall, and AI's usefulness to a student depends entirely on whether it is used to prompt that effort or to replace it.
Recognition is not retrieval
There is a trap worth naming before you build any of this. A multiple-choice question can often be answered by recognition, where the right option simply looks familiar, without the student ever recalling anything from a blank page. Recognition feels like knowing, which is exactly why students who coasted through a topic can still pick the correct box and leave convinced they understood it.
Real retrieval asks a student to produce the answer, not spot it. That is harder, and the difficulty is the point: within reason, the more effort the recall takes, the stronger the memory it builds, a finding researchers call desirable difficulty. So the useful classroom mix is not all one format. Fast option-based checks are good for reading the room in seconds. At least some prompts, though, should make students generate an answer from nothing, which is where short-answer questions and “explain it back” earn their place.
Where AI policy runs out
This is why questioning technique, not AI policy, may be higher education's most under-used lever. In the AAC&U survey, 87% of faculty said they had written explicit AI-use policies, yet 59% still said their institution was not well prepared to help students face what is coming.
Policy is necessary, but it governs what a student submits after the fact. It does little for the ninety minutes before they submit it, in the lecture hall or seminar room, where the actual thinking either happens or does not, and where no policy document has ever made anyone think harder.
What this looks like in a lecture
The practical move is small. It is a handful of retrieval questions dropped into the lecture itself, at the points where you would otherwise keep talking.
The Drexel study is worth copying almost exactly, because its design is the useful part: short prompts, during natural breaks, that ask students to recall or apply what they just heard rather than look it up. A live quiz is the cleanest way to run this with a full room, because it forces every student to commit to an answer instead of letting the confident few carry the session:
- اختر الإجابة for a fast recall or application check. “Which of these four assumptions does the model break if the data is not independent?” Everyone answers on their phones, the distribution appears on screen, and you teach straight into the wrong answers that show up.
- اجابة قصيرة when you do not want the options to do the remembering for the student. A typed prompt with no multiple-choice scaffold is closer to true retrieval, for the recognition reason above.
- الأسئلة التي تم إنشاؤها بواسطة الذكاء الاصطناعي, the way An and colleagues did it. The منشئ الاختبارات AhaSlides has an AI mode that drafts retrieval questions from a topic or an uploaded lecture file in a few seconds, which removes the main excuse for skipping the practice: that writing the questions is one more thing to prepare.
Three prompts worth stealing
The format matters less than what the prompt forces the student to do. Three that travel across disciplines:
- Free recall. “Without looking at your notes, write the one sentence that captures what the last fifteen minutes were about.” It is the cheapest retrieval question there is, and the blank sentences tell you instantly what did not land.
- Application to something new. “Here is a dataset you have not seen. Which method from today applies, and what breaks if you reach for the other one?” Applying a rule to a fresh case is far harder to fake than restating it.
- Spaced retrieval across sessions. “What did we cover last week that this week contradicts?” Pulling from an earlier class, not the last ten minutes, is where durable understanding gets built.
None of these take more than a minute of class time, and every one of them puts the effort back where AI cannot reach it.
The moments that still belong to the student
None of this is an argument for resisting AI in the classroom, and the data does not support that conclusion. Plenty of students already use it well.
It is an argument for being deliberate about which moments still belong to the student alone, and designing lectures, seminars, and assessments to protect those moments on purpose, one retrieval question at a time. One instructor in this edition has built an entire questioning practice around exactly that idea, replacing the reflexive “any questions?” with a question that makes students think. AI changed how fast a student can get an answer. It did not change what has to happen inside their head for the answer to become theirs.
مراجع حسابات
American Association of Colleges and Universities, & Elon University's Imagining the Digital Future Center. (2026, January 21). National survey: 95% of college faculty fear student overreliance on AI and diminished critical thinking among learners who use generative AI tools. AAC&U. https://www.aacu.org/newsroom/national-survey-95-of-college-faculty-fear-student-overreliance-on-ai-and-diminished-critical-thinking-among-learners-who-use-generative-ai-tools
An, Y., Liu, J., Acharya, N., & Hashmi, R. (2025). Enhancing student learning with LLM-generated retrieval practice questions: An empirical study in data science courses. arXiv. https://arxiv.org/abs/2507.05629
Stephenson, R., & Armstrong, C. (2026, March 12). Student generative artificial intelligence survey 2026 (HEPI Report 199). Higher Education Policy Institute. https://www.hepi.ac.uk/reports/student-generative-ai-survey-2026/
