Ask students directly and most of them will tell you AI is bad for their brains. Ask them what they did last night and there is a good chance the answer involves ChatGPT. That contradiction is not a talking point, it is a measured finding from a nationally representative survey, and it is the real story in education right now: usage and self-reported concern are both climbing at the same time, and the research on actual learning outcomes is split in ways that do not fit a tidy headline.
The numbers on student AI use for homework in 2026
RAND's American Youth Panel surveyed more than 1,000 students ages 12 to 29 and found that AI use for homework rose from 48% in May 2025 to 62% by December 2025. The increase was driven mostly by younger students: middle schoolers went from 30% to 46% using AI for schoolwork, high schoolers from 49% to 60%, while college student usage stayed roughly flat, likely because it was already high.
What are they actually doing with it? According to the same survey, the top uses are getting better explanations of assignments (38%), brainstorming ideas (35%), looking up facts (33%), and drafting or revising writing (33%). That is closer to how students already used Google and tutors than some of the cheating-focused coverage suggests, though drafting and revising written work is exactly the part that worries educators most.
Does AI make students worse learners? Here is the tension in the research
This is the real question, and the honest answer is: it depends heavily on how the AI is used, not whether it is used at all. Two studies from 2025 and 2026 point in opposite directions, and both are credible.
The first is MIT Media Lab's "Your Brain on ChatGPT" study, which had 54 Boston-area college students write essays while researchers tracked brain activity. Students who used ChatGPT showed lower neural connectivity, weaker memory recall of their own writing, and by the later sessions, many were copy-pasting rather than composing. The researchers called this pattern "cognitive debt." It is worth noting the paper has not been peer reviewed and the sample is small, but it lines up with something bigger: a 2026 RAND finding that 67% of students now say AI use harms critical thinking, up from 54% in early 2025. Faculty agree even more strongly. A College Board survey found 84% of faculty believe AI makes students more dependent on technology and less likely to think critically or engage with material, and a separate Elon University and AAC&U survey of faculty nationwide put that figure at 95%, with 48% saying they have already seen student research quality decline because of AI.
The second data point runs the opposite way. A Harvard study published in Nature Scientific Reports in 2025 gave 194 students a custom AI tutor built by physics lecturer Gregory Kestin, compared against an active-learning classroom, which is already considered a strong teaching method, not a passive lecture. Students using the AI tutor learned more than twice as much in less time and scored about 30% higher on a follow-up test. The difference was not the AI itself, it was how it was built: the tutor was instructed to give only one step at a time, never the full answer, and to keep responses to a few sentences to avoid overwhelming the student.
Why the same technology produces opposite results
Put the two studies side by side and a pattern emerges. The MIT study measured what happens when AI does the thinking for you: students asked ChatGPT to write, and it wrote. The Harvard study measured what happens when AI is deliberately restricted from doing the thinking for you: the tutor was built to withhold answers and force the student to work through each step. Same underlying technology, opposite design philosophy, opposite outcome.
This matches a broader pattern researchers have found in tutoring specifically. A NORC and Brookings review of AI tutoring research found that generative AI tutors combined with structured, Socratic-style prompting produce real learning gains, while unrestricted chatbot access tends to produce answer-copying instead of learning. The tool is not the variable. The constraint is.
The Khan Academy problem nobody expected
Even a well-designed AI tutor is useless if students do not use it. A two-year randomized trial of Khan Academy's Khanmigo, released by the National Bureau of Economic Research in August 2026, found it raised math achievement by about 1.26 national percentile ranks per school term, a real but modest gain. The bigger finding was that only 15% of students with free access to Khanmigo actually used it voluntarily. Khan Academy's own leadership described the tutor as "a non-event" for most students and had to redesign the product in 2026 to activate it automatically during practice sessions, because students were not seeking it out on their own.
That is a useful reality check on the entire debate. Before asking whether AI tutoring helps or hurts learning, the more basic problem is that most students, given the choice, do not use dedicated learning tools at all. They use whatever chat window is already open on their laptop, for whatever task is due that night, with no built-in restraint on how much of the work it does for them.
What this means if you are a student, parent, or teacher right now
- Using AI to get a concept explained differently or to check your own reasoning tends to track with the Harvard-style outcome: you still have to produce the answer yourself.
- Using AI to generate a finished draft and editing it lightly tracks with the MIT-style outcome: less retention, less ownership of the work, and by the study's own account, increasing laziness over repeated use.
- The line between those two uses is not about which app you open, it is about whether you do the cognitive work or the AI does it for you.
The takeaway
The honest answer to "does AI make students worse learners" is that AI itself is not the deciding factor, the design of the interaction is. A chatbot that hands over a finished essay measurably weakens the brain activity and memory tied to that task. A tutor engineered to withhold the answer and force step-by-step reasoning measurably improves test scores over a strong classroom alternative. Most students are currently using the first kind of tool for the second kind of task, which is exactly why two-thirds of them can tell you AI is hurting their thinking while still opening it every night. The technology is not the problem to solve. The habit is.
