I Learn, They Cheat: The Self-Serving Story Students Tell About AI

Cover Memo

This report draws on CHIP50 Wave 38.6 (total unweighted N = 23,462; fielded June 21–July 13, 2026), restricted to currently-enrolled students (school_recency = "currently enrolled"). Base sizes vary by item and are labeled throughout: the peer-perception and self-reliance items (ai_learn_others, ai_course_extent/extent2, ai_banned, ai_course_other) were asked of the full currently-enrolled sample (n ≈ 2,206–2,285); the self-learning and human-teacher items (ai_learn, ai_vs_tutor) were asked only of the ~1,430 currently-enrolled students who had used AI for schoolwork.

A note that is central to this report, not a footnote. Every figure here is a self-report. The report's core finding is a gap between how students describe their own AI use and how they describe their peers' — and a gap like that can arise because students flatter themselves, because they overestimate others, or both. We do not treat these as measures of actual behavior. The interesting fact is the asymmetry in students' own accounts, and where it is (and isn't) large.

All percentages and means are survey-weighted to national demographic targets; unweighted N is a reliability cue only. Descriptive differences are labeled as such; demographic claims are confirmed against survey-weighted regressions (Appendix). Partisanship uses the 7-point party7.

Key Takeaways

Introduction

The public conversation about students and AI is dominated by a single fear: that generative tools have become an engine of mass cheating. Students see the same headlines — and, it turns out, they largely apply the worry to other people. CHIP50's Wave 38.6 supplement asked currently-enrolled students both how AI affects their own coursework and how it affects their classmates', and the two accounts diverge sharply and consistently. Students describe their own use as a modest learning aid and their peers' use as heavy, and even harmful to learning.

This pattern has a name — pluralistic ignorance — and it is well documented for AI specifically. A University of Chicago study of undergraduates found that 60% admitted using AI themselves while 90% believed their peers did, a 30-point gap the authors attribute to some mix of under-reporting one's own use and over-estimating everyone else's.¹ It echoes decades of public-health research showing that false perceptions that "everyone else is doing it" can become self-fulfilling.² Because everything in this report is self-reported, that ambiguity is unavoidable — and it is precisely why the shape of the gap matters. As we show, CHIP50 students are fairly accurate about how many peers use AI; where their accounts diverge is in how much they think peers rely on it and whether it helps peers learn. That points to an evaluative asymmetry — a story students tell about who uses AI well and who abuses it — more than a simple miscount.

"I Learn; They Don't"

Students are far more generous about AI's effect on their own learning than their peers'. Among currently-enrolled students who use AI for schoolwork, 41% say it helps them learn more, 45% say it makes no difference, and just 14% say they learn less. But asked the same question about other students at their school, the optimism collapses: only 21% think AI helps their peers learn more, while 35% think it makes other students learn less — and another 15% say they aren't sure (Table 1).

Lined up, the asymmetry is stark. Students are roughly twice as likely to credit AI with helping their own learning as their peers' (41% vs. 21%), and two and a half times as likely to say it hurts others' learning as their own (35% vs. 14%). The generosity students extend to their own use — "it's a study aid that helps me understand" — is almost entirely withheld from everyone else, whose use they read as a crutch that erodes learning.

Comparative context. The self–peer split mirrors the University of Chicago finding that students see far more AI use in others than they admit in themselves.¹ CHIP50 adds that the gap is evaluative as well as quantitative: students don't just think peers use AI more, they think peers benefit less from it.

And They Say They're Not Coasting

A subtler version of the cheating fear is that even honest AI users quietly stop doing the hard cognitive work — that AI becomes a way to spend less time thinking. CHIP50 lets us test that directly. Asked how AI has changed the amount of time they spend studying or learning the material, currently-enrolled AI users split almost evenly into thirds: 33% say they now spend more time (12% "much more," 21% "somewhat more"), 34% about the same, and 32% less (25% "somewhat less," 8% "much less") (Table 6). The average lands squarely on "about the same."

That is not the picture the offloading worry predicts. If AI were mainly a shortcut, the responses would pile up at "less time"; instead the distribution is nearly symmetric, and about as many students say AI has drawn them into spending more time on the material as say it has let them spend less. It fits the self-learning story above: the 41% who say AI helps them learn more are, for the most part, describing a tool that adds engagement rather than subtracting effort. The usual caveat governs — "how much time I study" is exactly the kind of thing students may round in a flattering direction — but on their own accounts, the median AI user is not coasting.

"I Dabble; They're All In"

Students picture their peers as far more dependent on AI than they report being themselves — but only when it comes to intensity, not prevalence. Asked how much of their own schoolwork they do with AI, most currently-enrolled students describe light use: 24% say "none," 24% "not much," 34% "some," and only 14% "most" and 4% "all" — about 18% at the heavy end. Asked how much of their schoolwork other students do with AI, the distribution shifts dramatically toward heavy use: 44% believe their peers do "most" or "all" of their schoolwork with AI (Table 2). On a 1-to-5 "none-to-all" scale, students rate their peers' reliance (mean 3.3) nearly a full point above their own (2.5).

Yet this is not a case of students imagining AI everywhere. Asked to estimate what share of their classmates use AI at all, students put it at about 62% on average — almost exactly the 65% who actually report doing so (Table 3). In other words, students are close to right about how many peers use AI and far off about how much: they accept that most classmates use it, but imagine those classmates doing the bulk of their work with it, in a way few students say they do themselves.

The self-image is reinforced by what students admit about their own conduct. Nearly half — 46% — say they have used AI in ways that could be considered cheating at least once (25% once or twice, 15% several times, 6% frequently), leaving 54% who say never (Table 3). That is not a portrait of scrupulous abstainers; it is a portrait of students who cut corners themselves and still imagine their peers cutting more.

Who Leans In — and Who Still Prefers a Human Teacher

Beneath the self–peer story, the students most inclined to embrace AI are a familiar pair: Republicans and men. Survey-weighted regressions (Appendix) show that, controlling for the standard demographics, Republican and Strong-Republican students report the most cheating-adjacent use, are the most likely to say AI helps them learn, and rate AI most favorably against a human teacher; Democratic and independent students are consistently lower on all three. Men report more cheating-adjacent use than women and rate AI more favorably versus a teacher, though the gender gap on self-reported learning is not statistically reliable. Age, having been the illusory headline of the adoption story, works differently here: it does not predict how much AI students use (the flagship's central correction), but among current students it does predict how they judge it. Read on the survey's native age bands, older enrolled students are consistently more positive — every age band above 18–20 rates AI more favorably against a human teacher than the youngest students do, rising from +0.29 among 21–30s to +0.67 among 41–50s on the five-point scale — and students 31 and older are likelier to say AI helps them learn. On cheating-adjacent use, by contrast, there is no gradient at all: only the 51-and-older group stands apart, reporting markedly less of it.

That partisan tilt, on closer inspection, is a trust gap. Trust in AI companies is itself sharply partisan — Strong Republicans average 2.8 on a 1-to-4 scale, versus roughly 1.8–2.4 for everyone else (lowest among Democrats and independents) — and when trust in AI is added to the models, the partisan effect largely vanishes: party ID no longer predicts whether a student says AI helps them learn, and its effect on the teacher comparison is roughly halved, while trust becomes the single strongest predictor (doubling the model's explanatory power). Republican students embrace AI more, in short, in large part because they trust its makers more. Notably, this does not explain the gender or age patterns, which survive the trust control — only the partisan gap is a trust story.

Partisanship and gender are not the only independent axes. Race is a third, and it runs one way across the board. Net of party, gender, age, income, and the rest, Black students — and, on the teacher comparison, Hispanic students too — are more likely to say AI helps them learn, rate AI more favorably against a human teacher (Black +0.29, Hispanic +0.24 on the five-point scale), and report more cheating-adjacent use (Black +0.14, Hispanic +0.12). The favorable and unfavorable sides move together: this is not a group that is more virtuous or more suspect about AI, but one that is simply more engaged with it and says so. It is the same pattern the flagship finds for adoption and intensity and the affective report finds for excitement — Black students especially are among the most AI-embracing on campus once their partisan and economic composition is held constant.

There is a further twist that complicates the self-serving reading. The gap between how much students say they use AI and how much they believe their peers do is not spread evenly: it is smallest among Strong Republicans — who use AI heavily and say so — and largest among Lean Democrats and Democrats, who use it less but still picture their peers as all-in. Learning follows the same pattern: Republicans think their peers benefit from AI at rates close to how they rate themselves, whereas Democrats reserve the credit for themselves and are far more likely to say peers learn less. The self–peer asymmetry at the heart of this report is, in other words, disproportionately a left-of-center pattern.

For all this, students have not concluded that the machine beats the teacher. Among currently-enrolled AI users, only 29% rate AI as better than a human teacher for learning the material, 32% call it about the same, and 39% say it is worse (Table 4). The self-image that recurs across this report — I use AI in moderation, I learn from it, and I still value a real teacher, but my classmates overuse it and don't even learn — is real and consistent, but it is voiced most sharply by the students who use AI least and trust it least. Because it is self-reported, we cannot say how much of it is true; what we can say is who tells it.

Comparative context. That students keep a human teacher above AI is consistent with the Pew finding that students view AI as far more acceptable for research (54%) than for writing essays (18%) — an assistant, not a replacement.³

Conclusion

Ask students about their own AI use and they describe a disciplined study aid; ask them about their classmates and they describe an epidemic of dependence and cheating. The same students who say AI helps them learn think it makes others learn less; the same students who dab at it for a few tasks believe their peers do nearly everything with it; and the same students who admit their own corner-cutting impute heavier corner-cutting to everyone else. The one place the story holds together is prevalence: students know roughly how many of their peers use AI. Everything past that — how much, how well, how honestly — bends in a self-serving direction.

Because these are self-reports, the asymmetry cannot tell us who is right. It may be that students flatter themselves, that they overestimate their peers, or both. But the gap itself is the finding, and it carries a practical warning drawn from the same research on alcohol and drugs that first named pluralistic ignorance: when students believe "everyone else is doing all their work with AI," that belief can become permission. Institutions worried about AI and integrity may be fighting not only what students do, but what students wrongly believe everyone else is doing. What almost none of them are doing is settling the question in writing: Report 3 finds the classroom rules for AI largely absent, leaving the belief unchallenged.

Appendix A — Methods

Data. CHIP50 Wave 38.6, fielded June 21–July 13, 2026; total unweighted N = 23,462. Restricted to currently-enrolled students (school_recency = "currently enrolled").

Bases. ai_learn, ai_vs_tutor, ai_study_effort (asked of currently-enrolled AI users): n ≈ 1,356, 1,429, and 1,429. The three items sit behind the same route; ai_learn is 73 cases (about 5%) short of the other two purely through item-level non-response, not a narrower filter. No reweighting was applied for it, and the shortfall is spread across response categories rather than concentrated in one. ai_learn_others, ai_course_extent, ai_course_extent2, ai_banned, ai_course_other (asked of all currently-enrolled): n ≈ 2,206–2,285.

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_learn — "Do AI tools help you learn coursework material?" 3 = learn more · 2 = about the same · 1 = learn less (verified against the fielded questionnaire; the stored order is the reverse of an earlier data-dictionary draft). ai_learn_others — same about peers, 1 = others learn more … 3 = others learn less, 5 = not sure. ai_study_effort — "How has AI changed the amount of time you spend studying/learning the material?" 1 = much more time … 5 = much less time. ai_course_extent / ai_course_extent2 — self / peer share of schoolwork done with AI, 1 = None … 5 = All. ai_course_other — estimated % of peers who use AI (0–100 slider). ai_banned — "How often have you used AI in ways that could be considered cheating?" 1 = Never … 4 = Frequently. ai_vs_tutor — AI vs. a human teacher, 5 = AI much better … 1 = AI much worse. Partisanship: 7-point party7.

Self-report caveat. All estimates are self-reported and may reflect social-desirability bias; the self–peer gaps in particular cannot be decomposed into "students underreport their own use" vs. "students overestimate peers'."

Appendix B — Data tables

Table 1. AI and learning — self vs. peers (currently-enrolled; weighted %)

Self base ≈ 1,356 AI users; peer base ≈ 2,206 currently-enrolled students. "Other students" also includes 4% "they don't use AI." Weighted %.

Table 2. How much schoolwork is done with AI — self vs. perceived peers (currently-enrolled; weighted %)

Base ≈ 2,211 (self) / 2,285 (peers) currently-enrolled students. Scale verified 1 = None … 5 = All. Weighted %.

Table 3. Prevalence estimate and own conduct (currently-enrolled)

Base ≈ 2,209–2,214 currently-enrolled students. Weighted.

Table 4. AI vs. a human teacher (currently-enrolled AI users; weighted %)

Base ≈ 1,429 currently-enrolled AI users. Scale 5 = AI much better … 1 = AI much worse. Weighted %.

Table 5. What predicts students' AI attitudes (currently-enrolled; survey-weighted regressions)

OLS, survey-weighted, on party7 + gender + education + income + urban/rural + Census region + age band + race, among currently-enrolled students. Reference: Strong Republican, Female, College degree, Rural, Midwest, age 18–20, White. Age uses the survey's native age_cat_8 bins — 18–20 / 21–30 / 31–40 / 41–50 — with the three oldest bins collapsed into 51+ (age_stu5), because only 14 currently-enrolled respondents fall above age 60. n = 2,202 (ai_banned, all enrolled), 1,349 (ai_learn, users), 1,422 (ai_vs_tutor, users). ✓ = significant at p <.05. Adding trust in AI companies (pol_trust_ai) absorbs almost the entire partisan effect — party terms become non-significant on ai_learn and largely so on ai_vs_tutor, while trust becomes the strongest single predictor (see the report-2 trust/partisan-gap addendum). The gender and age effects are not trust-mediated; the race effects largely survive the trust control too, though the Black advantage on the learning item is partly absorbed by it. Urbanicity is null across all three outcomes; Census region is null on ai_banned and ai_learn, with one exception on ai_vs_tutor, where Southern students rate AI higher than Midwestern students (+0.19, p =.018).

Table 6. How AI changed study time (currently-enrolled AI users; weighted %)

Base ≈ 1,429 currently-enrolled AI users (ai_study_effort, 1 = much more … 5 = much less). Rows may not sum to exactly 100 due to rounding. Nets are computed on unrounded percentages and will not always equal the sum of the two rounded rows above them: "less time" is 24.85 + 7.63 = 32.5, displayed 32, even though the displayed components read 25 and 8. Weighted %.

Demographic robustness (SES, urbanicity, region)

The race findings are integrated in the body ("Who Leans In") and Table 5 (Black +0.11 on learning, +0.29 vs. teacher; Hispanic +0.24 vs. teacher; Black +0.14 / Hispanic +0.12 on cheating-adjacent use). On the other dimensions: lower education predicts more cheating-adjacent use; income effects are scattered. Urbanicity and Census region are null across all three outcomes — and the null is a narrow range, not an absence of analysis: student AI adoption runs from 59.1% in the Northeast to 67.9% in the West, with no region significantly different from another (model F p =.054).

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 … 5 = More than 10 years ago). Code 1 defines the currently-enrolled base.
• ai_learn — "Do AI tools help you learn coursework material?" (3 = Learn more; 2 = About the same; 1 = Learn less; 0 = Don't use AI for coursework). Note the descending code order: the stored codes run high-to-low, the reverse of an earlier data-dictionary draft.
• ai_learn_others — "Do AI tools help other students learn?" (1 = Learn more; 2 = About the same; 3 = Learn less; 4 = Don't use AI; 5 = Not sure).
• ai_course_extent — "How much of your schoolwork do you do with AI?" (5 = All; 4 = Most; 3 = Some; 2 = Not much; 1 = None).
• ai_course_extent2 — the same question asked about other students (same scale).
• ai_course_other_1 — "What percentage of other students use AI?" (0–100 numeric slider).
• ai_study_effort — "How has AI changed the amount of time you spend studying?" (1 = Much more; 2 = Somewhat more; 3 = About the same; 4 = Somewhat less; 5 = Much less).
• ai_banned — "How often have you used AI in ways that could be considered cheating?" (1 = Never; 2 = Once or twice; 3 = Several times; 4 = Frequently).
• ai_vs_tutor — AI compared with a human teacher (5 = Much better; 4 = Better; 3 = About the same; 2 = Worse; 1 = Much worse).
• pol_trust_ai — trust that AI companies "do what is right" (4 = A lot; 3 = Some; 2 = Not too much; 1 = Not at all), from the institutional-trust battery.
• party7 — derived 7-point party identification (1 = Strong Republican … 7 = Strong Democrat).