I have heard this argument before.
We have heard a lot about banning the use of AI in higher education lately. Plenty of faculty are struggling with the question of allowing or not allowing the use of AI in their classroom. These are reasonable people with real concerns. The shape of these questions and concerns is very familiar. We hear this question asked throughout our history. A decade ago I was in a room arguing about laptops and phones in class, and the assumption then was the same one operating now: that the amount of exposure is the thing to control.
We were wrong about that, and it took researchers eight years to show us why.
A team at the Universities of Jyvaskyla and Eastern Finland tracked 260 children through the PANIC study (Physical Activity and Nutrition in Children), following them from childhood into adolescence. The expectation going in was straightforward. More screen time, worse cognitive outcomes. That is not what the data showed. Teenagers who had accumulated more screen time as children showed better cognitive processing by adolescence.
I want to be careful here, because this finding gets misused. It is not a case for unlimited tablet time. Petri Jalanko, the study’s lead author, was specific about the condition: what matters is whether children use screens “in such ways that promote active thinking, problem-solving, creativity and learning.” Two kids with identical logged hours, one scrolling and one building something, were never having the same experience. The hours were the wrong unit of measurement.
Now swap “screen time” for “AI use.”
The Study Everyone Quotes
If you work in higher education, you have heard about MIT Media Lab’s “Your Brain on ChatGPT,” usually secondhand and usually in its most alarming form. The design was clean. 54 students were split into three groups and asked to write essays: one group unaided, one using a search engine, one using ChatGPT, all wearing EEG headsets. The brain-only group showed the deepest neural engagement. The search-engine group landed in the middle. The ChatGPT group showed the weakest brain connectivity, and it declined across sessions. 83 percent of that group could not accurately quote a line from the essay they had just submitted.
I take this seriously. I do not think it should be explained away, and I have stopped trying to soften it when faculty bring it up. But we should be precise about what it measured, which is what happens when a student hands the whole cognitive task to a chatbot and lets it run from prompt to finished paragraph. That is a specific behavior. That is not “AI in education. That is outsourcing using a shiny new tool. there.
Same Mistake, Different Technology
The Finnish researchers spent eight years discovering that “how much” is a question that mostly produces noise. What predicted outcomes was what the child was doing during those hours.
The AI literature is starting to say something similar. A 2025 study in Frontiers in Education compared ChatGPT with human tutors on critical thinking in university students, and the split result is the useful part. Students preferred the AI for low-judgment, self-paced exploration, but the interventions that actually moved critical thinking were the ones structured for dialogue and challenge rather than answer delivery. Randomized work on Socratic chatbots built to interrogate reasoning instead of resolving it reports gains in critical thinking and self-regulation.
The stronger evidence, at least to me, is a 2025 study in Scientific Reports of more than 800 engineering undergraduates. The authors traced a mediation chain: students who built genuine competence working with generative AI showed sharper critical thinking, which raised their AI self-efficacy, which predicted measurable gains in creativity. A companion study of graduate students found the same structure for research competence. AI helped, but only when it ran through the student’s own thinking rather than around it. The researchers’ summary is the line I keep repeating to our design team: “simply teaching technical skills is not enough.”
What this Looked Like on Our Campus
The abstract version of this argument has been made enough times. I lead the implementation of Socratic assistants in every course at Los Angeles Pacific University. We ran a pilot with a course level generative AI assistant called Spark. It engaged students in a Socratic way and had conversations with students about topics and assignments. It could not act like an answer vending machine. We did not set it loose. We spent time with faculty stress-testing it. It worked well. We added a course assistant to every course in the following term. The results were a rise in efficacy, motivation, and outcomes.
We then looked at outcomes and structured feedback in PSYC 105, comparing students who received AI-assisted instructor feedback with students who received instructor-only feedback, across 53 participants. The N is small but the data is clear. Students who got feedback naming a weakness and asking them to fix it engaged differently than students who got feedback that supplied the correction outright.
That is not a surprising result. It is the same result we have had for 40 years about human feedback. The part worth noticing is that the design principle transferred intact.
Abstinence is Not One of the Options
The question of whether students will use AI is settled. They are using it now, in study sessions, in problem sets, in the paper due at midnight, whether or not a syllabus permits it. And they are graduating into a labor market where employers assume competence with these tools on day one.
A prohibition does not protect students from the technology. It guarantees they learn it unsupervised, from each other, with nobody in the room who can explain the difference between outsourcing their thinking and sharpening it. I would rather be in that room.
Just like the screen time argument, it cuts in both directions. Nobody credible read the PANIC study and concluded that toddlers should get unlimited tablets. The takeaway was to teach the behavior that makes the time worth something. That obligation is heavier with a 20-year-old about to enter a competitive economy than with a child’s evening screen habits. Teach the behaviors students need in order to thrive and succeed today and in the future.
The Question we Keep Asking Wrong
“Should students use AI?” is our version of “should kids have screen time?” You can answer it, but the answer is useless, because it collapses a wide range of behaviors into a single variable. A student generating a finished essay and a student using a Socratic tutor to stress-test a thesis are not doing comparable things, any more than a child writing code and a child watching autoplay video are having comparable screen time. Lumping them together is how we end up with contradictory headlines and policies that ban a technology instead of teaching it.
The better question is narrower and harder to answer: does this particular use make the student do more cognitive work, or less?
Retrieval practice, argument construction, the discomfort of being asked “why” one more time than you would like. That is where learning happens, and it does not much matter whether the questioner is a professor, a teaching assistant, or a well-designed AI. An AI that routes around that work is a plagiarism machine with better manners. An AI that walks a student into it and will not let them leave early is doing something our faculty budgets have never been able to scale.
What I Check Before Adopting a Tool
If you are evaluating AI for a course or a program, the access dial is the wrong control. Here is what I look at instead, and it is a short list on purpose.
- Does it withhold? A tool that cannot be configured to refuse a direct answer will be used as an answer machine regardless of what the syllabus says.
- Does it ask before it tells? Check the default behavior on an ambiguous prompt. Does it request clarification, or does it produce an artifact?
- Can I see the reasoning, not just the output? If the transcript of the exchange is not available to the instructor, you have lost your only window into whether learning happened.
- Does it demand evidence? The useful behavior is pushing back on an unsupported claim, not polishing it.
- Is the effort visible in what the student submits? Design the assignment so the artifact captures the process. Otherwise you are grading the tool.
None of that requires a particular vendor, and most of it can be tested in 20 minutes with a trial account and one deliberately lazy prompt.
Where I’ve Landed
The MIT study is not the final word on AI and cognition. It is the first alarming headline about a new screen, and like the first screen time headlines, it documents a real failure mode without describing the whole system. The growing body of work on Socratic design and active learning suggests that the technology capable of eroding critical thinking is the same technology that builds it, and that the difference lies in how we configure and teach it.
We have had this argument before, about a different screen. What is different now is that sitting it out is not available to us. Our students are already using AI. What is still up to us is whether they learn to use it in a way that makes them better critical and creative thinkers than they would have without it.
