AI Goes to School: How Students Actually Use It

Cover Memo

This report draws on CHIP50 Wave 38.6, an AI/employment supplement to the COVID States Project's national survey, fielded June 21–July 13, 2026 (total unweighted N = 23,462).

Population. The AI-schoolwork battery was administered to a broad group that extended well beyond current students — including many respondents who last attended school years ago and whose adoption is far lower (16% among those who last attended >10 years ago, vs. 65% among the currently enrolled). Because the present-tense items ("how much of your schoolwork do you do with AI," "do your instructors encourage AI") only make sense for people actually in school, this report restricts to currently-enrolled students (school_recency = "currently enrolled"): unweighted n = 2,214 for the adoption item. Two sub-batteries are narrower and labeled where used: the frequency-of-use task battery was asked of the 1,431 currently-enrolled students who had used AI for schoolwork, and the intensity item of the ~2,211 currently-enrolled students.

Weighting. All percentages and means are survey-weighted to national demographic targets (race/ethnicity, age, gender, education, urban/rural); unweighted N is a reliability cue only. The student sample is boosted and not a probability sample of all U.S. students.

Self-report caveat. Every figure reflects what students report about their own AI use. Such self-reports may be self-serving — students may under-report uses they view as illegitimate (e.g., writing whole essays) and over-report "tutor-like" uses. Read the task-mix findings especially with that in mind.

Key Takeaways

Introduction

Generative AI reached students faster than almost any classroom technology before it. Pew Research Center found that the share of U.S. teens (ages 13–17) who use ChatGPT for schoolwork doubled in a single year — from 13% in 2023 to 26% in 2024.¹ Among college students, a 2026 Lumina Foundation–Gallup study reported that 57% use AI in their coursework at least weekly, with men using it daily at higher rates than women (27% vs. 17%).² The national picture is of a tool already woven into student life.

CHIP50's Wave 38.6 supplement lets us look more closely at how. It pairs a measure of whether students use AI for schoolwork with a battery on which tasks they use it for and how much of their work it touches. One measurement point matters throughout: the AI-schoolwork questions were asked of many people who are no longer in school — including many who last attended decades ago — and their much lower usage distorts any estimate of what students actually do. Restricting to currently-enrolled students — the population for whom "your schoolwork" and "your instructors" are live — changes the story, and this report is built on that restricted base. Throughout, remember that these are self-reports: measures of what students say they do, which may flatter the more legitimate uses.

The Topline: Two-Thirds Use AI, and Most Use It Lightly

Among currently-enrolled students, 65% have used an AI tool for their schoolwork. That is the headline number of this report, and it rests on a base narrowed to the people for whom the question is live — students in school now, rather than the much broader group the battery was administered to. On that broader group adoption is far lower: 16% among people who last attended school more than ten years ago. Two-thirds is therefore a statement about students, not about adults in general.

Using AI is not the same as leaning on it, and most students do not. Asked how much of their schoolwork they actually do with AI, about 24% of currently-enrolled students say "none," 24% "not much," 34% "some," and only 14% "most" and 4% "all" — roughly 18% combined. The distribution has a broad middle and thin tails: a quarter of students are outside the technology altogether, a little under a fifth have made it central to their coursework, and the majority in between reach for it on some tasks and not others. Adoption is close to a default; dependence is not.

Those two numbers — two-thirds in, fewer than one in five leaning heavily — are the frame for everything that follows. The next two sections describe what students reach for and what for; the section after that turns to the social divides in who uses AI and how much.

Comparative context. The 65% figure sits comfortably alongside the Lumina–Gallup finding that a majority of college students use AI in coursework — 57% at least weekly.²

ChatGPT Is the Default — and Little Else Comes Close

When students say they use "AI" for schoolwork, they overwhelmingly mean ChatGPT. Awareness of ChatGPT is essentially universal — 99.7% of all Wave 38.6 respondents have heard of it — and among currently-enrolled students it is not merely recognized but used regularly. On a scale from 1 ("never") to 7 ("several times a day"), students average 3.85 for ChatGPT — between monthly and weekly use, edging toward weekly — and somewhat higher for men (4.05) than women (3.70) (Table 5).

No other assistant comes close on frequency. Every other tool in the battery clusters well below ChatGPT, around occasional-to-monthly use or less, reached for only now and then by the smaller groups who have adopted it at all. (The tool-to-item crosswalk in this extract is positional and unverified, so only ChatGPT can be named with confidence; the "long tail" claim rests on the shape of the remaining means, not on identifying which assistant is which — see the appendix.) In practice the student "AI stack" is one dominant tool with a long, thin tail — which means that when institutions write AI policy, or when researchers track "student AI use," they are for now mostly describing behavior around a single product.

A Tutor, Not a Ghostwriter: What Students Use AI For

Among currently-enrolled students who use AI, the most common uses look more like tutoring than cheating. Asked how often they use AI for each of eight tasks (on a scale where 1 = "a lot" and 4 = "not at all," so lower means more frequent), students most often reach for it to explain a concept they didn't understand (mean 1.77) and to check their work for errors (1.86). Getting the answer to a specific problem (1.99) and generating a study guide or practice quiz (2.11) come next.

The task most associated with academic-integrity worries — writing, editing, or outlining an essay — ranks second-to-last of the eight (2.24), ahead of only scheduling and organizing schoolwork (2.54). In other words, the modal use of AI is comprehension and accuracy support, not wholesale outsourcing of written work. That said, this is precisely where the self-report caveat bites hardest: essay-writing is also the use students are most likely to view as illegitimate, so its low ranking may reflect what students are willing to report as much as what they do.

Gender differences are real but uneven. Men report more frequent use than women for getting answers, solving math, finding sources, and organizing schoolwork; the two are roughly even on the comprehension-oriented tasks (explaining concepts, checking work, making study guides).

Comparative context. The ordering matches how students judge acceptability: Pew finds teens far likelier to call AI acceptable for research (54%) than for essays (18%).¹

The Divides That Hold Up: Class, Party, Gender, and Race

The sharpest divide in student AI use is economic. Adoption climbs from 60% among lower-income students to 77% among upper-income students — a seventeen-point spread that survives controls for party, gender, education, urbanicity, race, and age. Income is the one demographic that clearly gates the door, and it is the access story taken up in report 4.

Beyond the door, the question is who leans on AI once inside — and whether the same students who adopt it are the ones who lean hardest. To test this, we modeled both adoption (whether a student uses AI at all) and intensity (how much of their work they do with it) on the same set of demographics, survey-weighted. Four patterns stand out:

• Partisanship predicts both, in the same direction. Republican students — especially Strong Republicans — are both more likely to use AI and more likely to do a large share of their work with it; independents and Democrats are lower on both. On a 1-to-5 "none-to-all" scale, Strong Republicans average 3.05, the highest of any group, versus about 2.3–2.4 for Democrats and independents. This is the strongest and most consistent demographic signal in the data, and it holds up under controls. (On adoption specifically, Republicans slightly outrank Strong Republicans — 79% vs. 75%; the Strong-Republican lead is on intensity.) What the party gap appears to be made of is trust, not ideology: Strong Republicans average 2.82 on the four-point AI-trust scale against 1.80 among Lean Democrats, and the partisan coefficients shrink substantially once that trust measure is added to the model (developed in report 6).
• Gender predicts both. Men adopt more and use AI more intensively than women (a modest but statistically reliable gap on both margins).
• Race is a third axis, and it points the same way. Net of everything else, Black, Hispanic, and Asian students all do a larger share of their schoolwork with AI, and Black and Asian students are also somewhat more likely to have adopted it at all. That adoption edge is a suppression effect — raw adoption is essentially equal for Black and Hispanic students (about 64%) and higher for Asian students (72%), and only once the lower average income of Black students is held constant does their adoption tick above comparable White students — but it is fragile: it holds with White as the reference category and weakens to non-significance once a White indicator is added to the model, so read the adoption result as suggestive. The intensity gap — all three groups do a larger share of their schoolwork with AI — is the robust one. Either way, Black and Hispanic students use AI for schoolwork at about the same rate as White students, and Asian students at a higher rate; adjusted for the income differences between these groups, their adoption rates are at least as high as those of comparable White students — a pattern developed across this series.
• Education and income open the door but don't deepen use. Higher prior education and higher income raise the odds of adopting AI — 71% of students who already hold a college degree and 67% with some college have used it, against 56% of those whose prior credential is a high-school diploma — but neither predicts intensity, and among users, more-educated students actually do slightly less of their work with AI (college degree −0.20 and some college −0.23 against a high-school-graduate reference).

Age is not one of these divides. Grouped into 18–20, 21–30, 31–40, and 41-and-over bands, adoption among current students runs 65%, 66%, 63%, and 67% — a four-point spread, with no contrast approaching significance either raw or under controls. Intensity is a partial exception: students aged 21–30 (+0.19 on the 1–5 scale) and 41 and over (+0.20) do modestly more of their work with AI than 18–20-year-olds, while 31–40-year-olds do not differ, so the pattern is a pair of small bumps rather than a gradient. The steep age slope visible in the raw data belongs to the population that was asked, not to students: it tracks how recently someone attended school, not how old they are. Age returns as a real divide the moment the subject changes from use to feeling: report 7 finds anxiety falling steeply with age across this same enrolled base.

Two cautions temper all of this. First, demographics are weak predictors here — the models leave most of the variation between students unexplained, so these are tendencies rather than deep cleavages. Second, these are self-reported intensities. Taken together, AI use among current students is broad, shallow for most, and only loosely patterned by who students are — with a real but modest partisan and gender tilt toward heavier use.

Conclusion

Among students actually in school, AI is close to a majority-default tool: two-thirds have used it. What varies is not so much who uses AI as how they use it, and even there the dominant pattern is restraint: most students use AI for particular tasks, lean on it lightly, and reach for it to understand material and check work more than to write their essays.

The divide that does not dissolve is income. Adoption climbs from 60% among lower-income students to 77% among upper-income students, and that gradient survives controls — the access story taken up in report 4. The other divides that survive are partisan, gender, and racial: Republican, male, and — net of income and party — Black, Hispanic, and Asian students all use AI more heavily. Age, by contrast, is not a divide among students at all: adoption sits between 63% and 67% from the youngest students to those past forty, and no age contrast is significant with or without controls. What there is not is a geographic story: adoption ranges only from 59.1% in the Northeast to 67.9% in the West, no region differs significantly from any other, and the pattern holds across urban, suburban, and rural communities alike. But even the partisan tilt is a modest one on top of a much larger shared reality, and everything here rests on what students are willing to say about themselves. This report establishes the landscape — broad, shallow, and patterned mainly by class and party. The reports that follow ask the sharper questions it raises: why students suspect their peers of the cheating they describe without hesitation, where the classroom rules are, and for whom AI is an equalizer.

Appendix A — Methods

Data. CHIP50 Wave 38.6, fielded June 21–July 13, 2026; total unweighted N = 23,462.

Population & base. Restricted to currently-enrolled students (school_recency = "currently enrolled"). Adoption item (ai_course): n = 2,214. Intensity item (ai_course_extent): n ≈ 2,211. Task battery (ai_school_use_1–8): n = 1,431 currently-enrolled AI users. The student enrollment-type item is not available in this data extract; school_recency is used to define current enrollment.

Weighting. Survey-weighted to national demographic targets; unweighted N is a reliability cue only. Not a probability sample of U.S. students.

Measures (verified coding). ai_course (1 = Yes used AI for schoolwork, 2 = No, 3 = Not sure); adoption = % Yes. ai_school_use_1–8 (1 = A lot … 4 = Not at all; lower mean = more frequent). ai_course_extent — "How much of your schoolwork do you do with AI?" 1 = None … 5 = All (verified against the fielded questionnaire; the stored order is the reverse of an earlier data-dictionary draft). ai_know_1 — heard of ChatGPT (Yes/No; 99.7% Yes). ai_freq_x1 — ChatGPT use frequency (1 = never … 7 = several times a day), asked of those who have heard of it. Partisanship uses the 7-point party7 (1 = Strong Republican … 7 = Strong Democrat).

Modeling. Adoption modeled by survey-weighted logistic regression, intensity by survey-weighted OLS, each on party7, gender, education, income, urban/rural, the multi-select race binaries, and age band, among currently-enrolled students. Age enters as four bands collapsed from the survey's native age_cat_8 bins — 18–20 / 21–30 / 31–40 / 41 and over (recode age_stu4; unweighted n = 759 / 779 / 366 / 306) — because the currently-enrolled sample thins sharply above 40: of the 306 enrolled respondents aged 41 or older, only 14 are above 60 and one is above 70. Standard errors are model-based, not design-based. A difference is described as reliable only where its coefficient is significant at p <.05.

Self-reports. All estimates are self-reported and may be subject to social-desirability bias, especially on integrity-adjacent uses.

Appendix B — Data tables

Table 1. Ever used AI for schoolwork, currently-enrolled students (weighted %)

Base: 2,214 currently-enrolled students (2,210 on the age item). Weighted %. Wave 38.6. Age rows use four bands collapsed from the survey's native age_cat_8 bins — 18–20 / 21–30 / 31–40 / 41 and over, unweighted n = 759 / 779 / 366 / 306 — and no age contrast is significant, adjusted or unadjusted.

Table 2. What students use AI for (currently-enrolled AI users; weighted mean, 1 = a lot … 4 = not at all)

Base: 1,431 currently-enrolled AI users. Lower mean = more frequent use. Wave 38.6.

Table 3. How much of their schoolwork students do with AI (currently-enrolled; weighted %)

Base: ~2,211 currently-enrolled students. Scale verified 1 = None … 5 = All. Rows may not sum to 100 due to rounding. Wave 38.6.

Table 4. What predicts adoption vs. intensity (currently-enrolled; survey-weighted models)

Adoption: weighted logistic on ai_course=Yes (n = 2,203). Intensity: weighted OLS on ai_course_extent, all currently-enrolled (n = 2,204) and users only (n = 1,426). Reference group: Strong Republican, White, Female, HS graduate, lowest income, Rural, age 18–20. Age enters as four bands collapsed from the native age_cat_8 bins: 18–20 / 21–30 / 31–40 / 41 and over. ✓ = significant at p <.05. Demographics account for only a small share of the variation in either outcome.

Table 5. ChatGPT use frequency among currently-enrolled students (weighted mean, 1–7)

Base ≈ 1,987 currently-enrolled students who have heard of ChatGPT (ai_freq; 1 = never, 3 = about once a month, 4 = about once a week, 7 = several times a day). ChatGPT awareness is 99.7% of all respondents (ai_know). ChatGPT is the most-used AI tool among students by a wide margin; reliable per-tool figures for the other assistants are not available in this data extract. Weighted.

Demographic robustness (SES, urbanicity, region)

The race findings are in the body (see "the divides that hold up"). On the other dimensions: SES is real and central — higher income and education raise adoption (the access story), though they do not deepen intensity. Urbanicity and Census region are null for adoption and for intensity among all enrolled students (suburban +0.13, p =.082) — the raw regional adoption range is a narrow 59.1% (Northeast) to 67.9% (West), with no region significantly different from another (model F p =.054); the one exception is a modest suburban edge on intensity among AI users specifically (+0.18, p <.05). The race intensity effects (Black +0.34, Hispanic +0.14, Asian +0.22, all p<.05) are robust; the race adoption ORs (Black 1.26, Asian 1.45) are reference-dependent — significant with White as the implicit reference but not once race_white is added, with Black near the p=.05 line — so treat the adoption suppression as suggestive, not firm.

Appendix C — Question wording and response codes

Source: the Wave 38.6 Qualtrics instrument provided by the CHIP50 / COVID States team (the Wave 38.6 codebook, 2026-07-20), which is the authoritative record of item wording and stored numeric codes for this wave. Stems below are the codebook's condensed forms, not screen-verbatim text; response options and codes are verbatim. Where an item's stored code order differs from its on-screen display order, the stored codes are authoritative and are the ones used throughout this report.

• school_recency — "When did you last attend school or university?" (1 = Currently enrolled; 2 = Within the past 2 years; 3 = 3–5 years ago; 4 = 6–10 years ago; 5 = More than 10 years ago). Code 1 defines the currently-enrolled base used throughout.
• ai_course — "Have you ever used AI for schoolwork?" (1 = Yes; 2 = No; 3 = Not sure). Adoption = % Yes.
• ai_school_use_1–8 — "How often do you use AI for each of the following?" (1 = A lot; 2 = Some; 3 = Not much; 4 = Not at all). Items: 1 explain a concept · 2 check work for errors · 3 generate a study guide or quiz · 4 get the answer to a problem · 5 write/edit/outline an essay · 6 find sources or citations · 7 solve math · 8 schedule or organize schoolwork. Lower mean = more frequent use.
• ai_course_extent — "How much of your schoolwork do you do with AI?" (5 = All; 4 = Most; 3 = Some; 2 = Not much; 1 = None). Reported here on the ascending 1 = None … 5 = All reading.
• ai_know_1 — prior awareness of the first-listed tool in the 18-tool AI-awareness battery; a binary awareness indicator, 99.7% aware. Value labels for this battery could not be confirmed against the Wave 38.6 codebook, so no response codes are printed here, and the tool index is positional in this extract with only x1 (ChatGPT) confirmed — see Appendix A.
• ai_freq_x1 — "How often do you use ChatGPT?" (1 = Never; 2 = Once or twice; 3 = About monthly; 4 = About weekly; 5 = Multiple times a week; 6 = Once or twice a day; 7 = Several times a day). Asked of those who have heard of the tool. The ai_freq_x{n} tool index is positional in this extract and only x1 (ChatGPT) is confirmed — see Appendix A.
• party7 — derived 7-point party identification (1 = Strong Republican … 7 = Strong Democrat).