For college students AI is now routine: in Gallup's 2026 survey of 3,801 students, 57 percent use AI for coursework at least weekly and about 20 percent daily, most often for help with coursework (64 percent) and checking answers (60 percent). Those are the uses where the tool does the step. In the only randomised trial to test that behaviour, Bastani and colleagues (2025) found a standard chatbot raised practice scores 48 percent and cut unassisted exam scores 17 percent, while a version prompted to give hints without answers raised practice scores 127 percent and left exam scores unchanged. Students in the harmed group did not notice. Two thirds of students told RAND in December 2025 that AI use harms critical thinking, and below-median-GPA students adopt fastest. The test that separates the two kinds of use is whether you can rebuild the reasoning with the tab closed.
On this page+
For college students AI is now routine. In Gallup's 2026 survey, 57 percent said they use AI for coursework at least weekly and about one in five use it daily. The two most common uses are help with coursework and checking answers. In the only randomised trial to test that kind of use, it raised practice scores and lowered exam scores.
You are probably a student, or close to one. There is a chat window open next to the problem set, the tool explains the step you were stuck on, the assignment goes in on time, and a small voice asks whether you could have done that step yourself. Every article on this question is written for your dean. This one is written for you, and it comes down to a single test you can run tonight.
For college students AI is routine: how much, and for what?
More than half of college students use AI every week, and the uses are mostly the ordinary work of getting through a course. The Lumina Foundation and Gallup surveyed 3,801 associate and bachelor's students in October 2025 and published the results in April 2026. Fifty-seven percent use AI in their coursework at least weekly; about 20 percent use it daily, with men (27 percent) ahead of women (17 percent). Asked what they use it for at least weekly, students said help with coursework (64 percent), checking homework answers (60 percent), editing their writing (54 percent), summarizing lectures or notes (54 percent), generating ideas (49 percent), research (45 percent), writing papers (36 percent) and career advice (35 percent). The reason they give most often, at 89 percent, is understanding complex material.
Campus policy has not kept up: 42 percent say their school discourages AI, 11 percent that it prohibits it, and 35 percent that it encourages it with limits.1 The pattern holds inside the most selective schools, where AI in higher education arrived fastest. Zara Contractor and Germán Reyes surveyed 634 students at Middlebury College between December 2024 and February 2025 and found over 80 percent using AI for academic work, up from under 10 percent two years earlier. Their top uses match Gallup's: explaining concepts (80 percent), summarizing texts (74 percent), finding information (63 percent), generating ideas (62 percent), and, at 24 percent, writing whole essays.2
Read those two lists again and notice what they have in common. Almost every entry is the tool doing a step the course was designed to make you do.
Which uses help you learn, and which only feel like it?
The uses of AI for studying that help are the ones where you still do the step; the uses that only feel like help are the ones where the tool does it and you read the result. Contractor and Reyes split student use into two kinds. "Augmentation" is when the AI works with the student: an explanation, feedback on a draft. "Automation" is when the AI does the work for the student: the summary, the paragraph, the solved problem. Among users, 56 percent reported augmentation and 45 percent automation, and most students do both.2
The split by name is less useful than the split by what you can rebuild afterwards. An explanation you read and then reproduce on paper was augmentation. The same explanation, read and nodded at, was automation with a friendlier name. The tool cannot tell which one happened. Only you can, by trying the step again without it.
What you asked for
Who does the step
What you can rebuild after
Holds up on exam day
The final answer, to check yours
The tool, unless yours was already written
Whether you were right, not why
Rarely
A worked solution to a problem you were stuck on
The tool
The shape of the solution, for a day
Rarely, unless you redo it closed-book
A hint on the next step only
You, with a nudge
The step, and the one after it
Usually
A summary of the reading
The tool
The headings
No; the exam asks about the middle
Feedback on a draft you wrote
You wrote it; the tool marks it
Your own argument, improved
Yes
An explanation of a concept, then you explain it back
You, on the second pass
The concept
Yes
Two things about the Gallup list now stand out. "Checking answers" is the cheapest use of all when your answer was already on the page and the most expensive when it was not. And "help with coursework" is not a use. It is a bucket, and which row of the table it lands in is decided by the question you type.
What happens on exam day when the tool is gone?
In the one randomised trial that measured it, students who used a standard chatbot during practice scored 17 percent worse on the unassisted exam than students who used nothing, and they did not notice. Hamsa Bastani and colleagues at Wharton ran the trial with nearly 1,000 high school students in Turkey in autumn 2023, across four 90-minute maths sessions covering about 15 percent of the term's curriculum. Students were assigned to three groups: textbook and notes only, a standard chatbot interface the authors call GPT Base, or GPT Tutor, the same model prompted to give hints without the answer and loaded with the teachers' own solutions and common mistakes. In the practice sessions both tools helped a great deal: scores rose 48 percent with the basic chatbot and 127 percent with the tutor. Then the tools were taken away for the exam. The tutor group scored the same as the control group. The basic chatbot group scored 17 percent lower.3
The mechanism is the one the Gallup list predicts. The transcripts showed that students with the basic chatbot "most often simply ask for the answer". The authors could also tell copying from misunderstanding: arithmetic errors and logical errors in the chatbot's output hurt practice scores by about the same amount, and neither carried into the exam, which is what you would expect if answers were being transcribed rather than read.3 And the students felt fine. Asked afterwards, the basic-chatbot group did not believe they had performed worse or learned less.
A tool that hands you the step feels like learning. A tool that asks you for the step is learning.
Two honest caveats. These were teenagers doing algebra, not undergraduates writing essays, and the effect was measured over a few weeks. But the behaviour under test, asking the tool for the answer while you are stuck, is the behaviour 60 percent of college students report weekly. We have written up why handing over the thinking costs the skill in full; here the point is narrower. The same tool produced both results. What changed was whether it was allowed to give the step away.
Why do the students using AI most worry most?
Because they are the ones watching it happen. RAND's American Youth Panel asked students from middle school through college, in February and again in December 2025, whether they agreed that "the more students use AI for their schoolwork, the more it will harm their critical thinking skills". Agreement rose from 54 percent to 67 percent in ten months, while using AI for homework rose from 48 percent to 62 percent.4 Among college students who use AI for homework, 64 percent agree that it harms the skill they are paying to build. Across all ages, non-users agree more (78 percent) than users (60 percent), but a majority of users saying it about their own habit is the number worth sitting with.
It is the same pattern we found in the school-age data, where the cons of AI in education, ranked by evidence came with the twist that students who used AI more were more concerned than adults who did. Contractor and Reyes add a detail that turns over-reliance on AI from a worry into a distribution: students below the median GPA adopted AI faster (87 percent) than students above it (80 percent), so the students with the most to lose are the ones using it most.2 And guidance is thin. Fifty-seven percent of college students told RAND that some of their teachers allow AI help and some do not, so the rule changes with the room.4
In sessions we see the same unease from people well past college: a fluent answer they accepted last week and cannot now reproduce. The unease is accurate. It is what the exam-day result feels like from the inside, and it is the AI and critical thinking question in one sentence.
How do you use AI in college without losing the skill?
Use it in the ways where you still do the step, and check yourself with one rule: if you cannot rebuild the reasoning with the tab closed, the tool did that part and you did not learn it. Three moves make that the default rather than the exception.
First, ask for the hint, not the step. Bastani's tutor was the same model as the chatbot that harmed learning; the difference was a prompt that withheld the answer.3 You can write that prompt yourself: "Do not give me the answer or the next step. Ask me what I would try first." If the tool gives you the step anyway, that is the moment to stop reading.
Second, write before you read. Put your own attempt on paper before you open the chat, even a wrong one. Then the tool's explanation has something to land on, and the gap between the two is the learning. This is the smallest form of how critical thinking actually improves: one real problem, on paper, with a question.
Third, close the tab and redo it. Within the hour, while the explanation is warm. If it comes back, it is yours.
The exercise is what a session at Inwitt AI is built around. It is an AI thinking partner app that does not give the step: it asks one question at a time, your reasoning goes onto a thinking board in your own words, and when you ask for facts the guess-first research reveal makes you commit to a guess before it shows what it found. The tool holds the question. You do the step.
Common mistakes college students make with AI
The first mistake is treating "it explained it" as "I learned it". Understanding an explanation and producing one are different skills, and the exam tests the second. Contractor and Reyes found 68.5 percent of students believe AI improves their understanding of course material; Bastani found the students whose scores fell believed the same.23 Both can be true. You understood it at the time.
The second is using one prompt for two jobs. Checking an answer you already wrote and getting an answer you had not written are the same request to the tool and opposite events for you. Write the answer first, or do not ask.
The third is judging by the grade on the assisted work. The practice scores in the trial went up 48 percent, and that number measured nothing the students kept. The only grade that tells you what you learned is the one on work the tool did not touch.
The fourth is waiting for the syllabus to decide for you. With 57 percent of students reporting rules that change by teacher, the policy on AI in university courses will not settle the question of what you can rebuild.4 The steps with a written output are the same whether the tool is allowed or not.
Questions people ask about college students and AI
Is using AI for homework cheating?
It depends on the use and the course, and students draw the line where the learning stops. In RAND's December 2025 survey, 78 percent said asking AI for a better explanation of an assignment was not cheating and 72 percent said the same about brainstorming, while a sizable minority called getting the answers or the writing done cheating.4 The rebuild test is a fair proxy for the line: if you could not redo it, someone else did it.
Do colleges allow AI?
Mostly with conditions, and the conditions vary by course. Gallup found 42 percent of students say their school discourages AI, 11 percent that it is prohibited, 35 percent that it is encouraged with limits and 7 percent that it is encouraged freely.1 Contractor and Reyes found that an explicit prohibition cuts intended use from 53 percent to 13 percent, so the rule does change behaviour where it exists.2
Is AI making students worse at critical thinking?
The students think so, and the one trial that measured skill after use found a 17 percent drop for the basic chatbot and none for the tutor that withheld answers.34 The honest answer is that it depends on which of the two uses you make of it, and the tool does not choose that for you.
Which students use AI the most?
Men more than women, business, technology and engineering students more than others, and, in the elite-college data, students below the median GPA more than those above it.12 Adoption is highest where the coursework is most solvable by a tool, which is where the exam-day gap is easiest to hide.
Every session leaves a board and a receipt, in your own words, of the steps you produced yourself. If there is a course, a problem set, or a habit you want to think through before the tool does it for you, join the Inwitt waitlist and bring it.
Contractor, Z. and Reyes, G. (2025, revised April 2026). "Generative AI in Higher Education: Evidence from an Elite College." arXiv:2508.00717. Survey of 634 Middlebury College students, December 2024 to February 2025. https://arxiv.org/abs/2508.00717↩↩2↩3↩4↩5↩6
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., Mariman, R. (2025). "Generative AI without guardrails can harm learning: Evidence from high school mathematics." Proceedings of the National Academy of Sciences 122(26). Randomised trial with nearly 1,000 students in grades 9 to 11, autumn 2023. https://pmc.ncbi.nlm.nih.gov/articles/PMC12232635/↩↩2↩3↩4↩5
Schwartz, H. L. and Diliberti, M. K. (2026). "More Students Use AI for Homework, and More Believe It Harms Critical Thinking: Selected Findings from the American Youth Panel." RAND Corporation, RR-A4742-1, 17 March 2026. Surveys of 2,039 students in May 2025 and 1,214 in December 2025, of whom 292 were in college. https://www.rand.org/pubs/research_reports/RRA4742-1.html↩↩2↩3↩4↩5
