Students did not wait for schools to finish the AI rulebook.
Between May and December 2025, the share of American students from middle school through college who said they used AI help for homework rose from 48 to 62 per cent. Over the same year, concern about the effect of these tools grew too. By December, 67 per cent agreed that the more students use AI for schoolwork, the more it will harm their critical-thinking skills.
Only about one-third said their school had a schoolwide rule governing AI help with homework.
Those figures create a tempting story about hypocrisy: young people fear AI, use it anyway, and receive little guidance from adults. The fuller picture is more interesting. “Using AI” covered everything from asking for a clearer explanation to requesting a homework answer. Students judged those activities differently, and policies often changed from one classroom to the next.
The findings come from a RAND report published in March 2026. It offers a useful national snapshot, but it is a survey of behaviour and belief. It does not show that AI caused a measurable decline in anyone’s reasoning.
The headline compares national snapshots
RAND’s main December survey included 1,214 enrolled Americans aged 12 to 29. It was in the field from December 4, 2025 to January 5, 2026. The sample contained 184 middle school students in grades 5 to 8, 738 high school students in grades 9 to 12, and 292 undergraduates.
Participants came from traditional public, charter and private schools, homeschools and postsecondary institutions. RAND weighted the responses to represent enrolled US youth with respect to gender, race or ethnicity, grade level and region.
There is an important wrinkle in the May-to-December comparison. The May survey had 2,039 eligible respondents and the December survey had 1,214. The samples were independently drawn and compared cross-sectionally, even though 848 December respondents had also completed the May survey.
In other words, RAND did not calculate how the same named students changed over seven months. It compared two weighted snapshots. Eligibility expanded from ages 12 to 21 in May to ages 12 to 29 in December, and the way college students were sampled changed. The researchers did not run significance tests across the survey waves and urged caution with the college trend.
That does not make the 48-to-62 movement meaningless. It means the cleanest conclusion is that reported use was substantially more common in the later national snapshot, especially among school-aged students.
Middle and high school drove the rise
Among middle school students, the share reporting AI help for homework rose from 30 per cent in May to 46 per cent in December. Among high school students, it rose from 49 to 63 per cent.
College students began far ahead and barely moved, from 71 to 72 per cent. Because of the sampling changes, even that apparent stability should not be treated as a precise measure of what happened across all US colleges.
December’s survey also asked about seven categories of AI tools. Seventy-one per cent used at least one. Chatbots were most common at 60 per cent, followed by AI writing helpers at 21 per cent and general homework helpers at 15 per cent. Tools designed specifically for tutoring were used by 8 per cent.
ChatGPT remained the dominant named product. Fifty-three per cent said they used it in class or for homework, up from 45 per cent in May. Google Gemini rose from 13 to 28 per cent. Those product figures describe use, not frequency, depth or whether the output was submitted.
The broad adoption fits an earlier pattern we examined at Silicon Canals: AI assistance often helps less-experienced people complete tasks faster. But education is not a call centre. In school, completing the task is only part of the objective. The difficult mental work may be the product teachers are trying to develop.
AI use covered several different acts
When RAND asked why students used AI for school, no single purpose dominated. Thirty-eight per cent used it to get better explanations of assignments. Thirty-five per cent used it for brainstorming. Looking up facts and drafting or revising writing each reached 33 per cent.
Students typically selected two of the eight listed purposes. In written responses, some described checking answers, understanding where they had gone wrong and asking a system to walk them through a problem.
These are not educationally interchangeable. A student who asks, “Why does this equation need that step?” is doing something different from a student who pastes in the problem and copies the result. Drafting a complete essay is different again from asking for objections to an argument the student has already made.
That variety matters when interpreting any single measure of AI use. It also helps explain why students could be frequent users without viewing every use as legitimate.
Students drew their own line around cheating
Among students who used AI to get better explanations, 78 per cent said that use was not cheating. For brainstorming, 72 per cent said it was not. Sixty-seven per cent said the same about looking up facts. Writing prompted more hesitation than any of those three uses.
Homework answers were the exception. Of the students who used AI for that purpose, 45 per cent called it cheating. Many of the others said the judgment depended on the situation. The uncertainty was much greater than it was for explanations or brainstorming.
There is another limit. RAND asked the cheating question only of students who said they used AI for that specific purpose. Students who rejected a use because they considered it cheating never received the follow-up. The resulting percentages may therefore understate how widely a use is considered improper.
Even so, the pattern is coherent. Students were more permissive when AI clarified, prompted or supplied background. They were far more divided when it supplied the thing the assignment appeared to ask the student to produce.
The 67 per cent figure measures concern, not lost ability
The critical-thinking question asked students how strongly they agreed with this claim: “The more students use AI (such as ChatGPT) for their schoolwork, the more it will harm their critical thinking skills.” Fifty-four per cent agreed or strongly agreed in February 2025. Ten months later, 67 per cent did.
Concern rose at every educational stage and rose most sharply among middle school students. Yet use continued to spread.
Even among December’s homework AI users, 60 per cent believed heavier use harms critical thinking. Among non-users, 78 per cent believed it. Within the user group, the figures were 52 per cent for middle school students, 60 per cent for high school students and 64 per cent for college students, though RAND said those grade differences were not statistically significant.
This is evidence about perception. The survey did not administer a reasoning test, inspect student work or randomly assign AI access. It cannot tell us whether a particular use improves or weakens learning, or whether worried students use AI differently from unworried ones.
There is no need to dismiss the concern simply because it is self-reported. Students may recognise a change in the amount of effort they invest. But recognising a plausible mechanism is not the same as measuring its effect.
The contradiction is smaller than it first appears
A student can believe that overreliance is harmful and still find a targeted use helpful. Most of us understand that a calculator can save time without concluding it belongs in every stage of learning arithmetic. The educational question is not whether an instrument exists, but when the work it removes is work the learner still needs to practise.
Our earlier look at the largest human-AI creativity comparison yet reached a related boundary. An AI system can produce a strong answer on a task without establishing what a person learned while using it. Output quality and skill formation are separate outcomes.
The students’ responses may therefore be less a confession of hypocrisy than an accurate description of an unresolved trade-off. AI is quick, available and sometimes genuinely explanatory. It can also make it easy to bypass the struggle through which understanding is built.
Most students did not report one shared rule
About one-third of middle school, high school and college students said their school had a schoolwide rule about AI help for homework. Where a shared rule existed, respondents were two to three times more likely to say it prohibited AI help than allowed it.
For many others, the answer depended on the teacher. That was reported by roughly 34 per cent of middle school students, 44 per cent of high school students and 57 per cent of college students. Some of the remainder said there was no rule or that they did not know.
This was not an audit of written policies. It was a measure of what students understood their school’s rule to be. A policy hidden in a handbook may count on paper while functioning like no policy at all. Conversely, a teacher may give precise assignment-level instructions even when no schoolwide rule exists.
Discretion is not automatically a flaw. What is acceptable in a language exercise may be inappropriate in a take-home exam. A blanket rule can be too blunt. The problem comes when the variation is invisible until a student is penalised.
RAND’s earlier national work on AI guidance in schools had already found that use was outpacing clear direction. The December findings show the gap from the student side: teachers may be making local decisions, while many learners still lack a stable baseline.
A useful policy asks who is doing the thinking
RAND proposes a more useful distinction than “AI” versus “no AI.” Cognitive offloading happens when the system does the mental work for the student. Cognitive augmentation happens when the system prompts the student to do deeper or more careful work.
The boundary is not perfectly clean, but it can guide concrete rules. Asking for a final answer and submitting it is mostly offloading. Asking for a counterargument, explaining why an answer is wrong, comparing two approaches or receiving feedback on reasoning can preserve more of the work.
Good guidance should operate at two levels. A schoolwide baseline can define disclosure, privacy, authorship and unacceptable substitution. Teachers can then explain which uses fit a particular task, why they fit, and what evidence of independent thinking students must show.
The same prompt may be sensible in one phase and damaging in another. AI feedback on a completed draft may help a student revise. Generating the first draft before the student has formed an argument may remove the very practice being assessed.
Policy also needs language a 12-year-old can apply at home. “Use AI responsibly” leaves the hard decision untouched. “You may ask for explanations and practice questions, but you may not ask it to write your response; include the prompts you used” gives a learner something operational.
The schools are already late, but not powerless
The report does not justify panic. It also does not support waiting for a perfect body of evidence while students make the rules privately.
By December 2025, AI help with homework was ordinary behaviour for a majority of enrolled young Americans in RAND’s sample. Many of those same students were uneasy about the effect on their thinking. Their answers suggest they can already see differences between explanation, collaboration and substitution, even when institutional rules lag behind.
The most useful response is neither a universal ban nor a vague endorsement. It is a clear account of what a given assignment is meant to teach, which forms of assistance preserve that goal, and which ones quietly erase it.
Schools are no longer deciding whether AI enters homework. It already has. They are deciding whether students will meet it through an intelligible set of expectations or through a patchwork of guesses, habits and penalties. The 62 per cent adoption rate shows the speed of the change. The 67 per cent concern rate shows that students understand there is something worth protecting.
