Engaged, Uneasy, and Uncritical: Using AI More Makes Students More Excited, Not More Careful

Cover Memo

This capstone synthesizes the eight-report "AI Goes to School" series, all drawn from 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). Its primary population is currently-enrolled students (school_recency = "currently enrolled", unweighted n ≈ 2,206–2,332 depending on the item), the base used throughout the series; Report 5's degree-value, high-school, and parent findings draw on the broader adult, high-school, and parent samples and are labeled where cited.

All figures are survey-weighted to national demographic targets; unweighted N is a reliability cue only. Every number is a self-report — what students say about their use, feelings, and knowledge, not audited behavior or measured learning; the reports treat the self–peer and self-assessment gaps as findings in their own right rather than as ground truth.

This report collects no new data. It rests on the series' already-verified figures (each independently re-derived and reviewed) plus one new integrative analysis run for the capstone: how students' feelings and AI-reliability knowledge vary with how much of their schoolwork they do with AI (Tables 1–2). Its purpose is to state, and test, the thesis the eight reports imply — that within the student body, the students who have gone furthest with AI are no more at ease with it and no more critical of it than those who have barely started.

The argument in one line

Read on their own, the eight reports in this series are eight findings. Read together, they are one: going further with AI makes students more enthusiastic about it without making them more comfortable or more careful. Three things are simultaneously true of the same students — they are deeply engaged with AI, genuinely uneasy about it, and largely uncritical of what it hands them — and the surprising part is that these do not cancel out. Using AI more does not resolve the unease, and it does not sharpen the scrutiny. It just adds enthusiasm on top.

1. Engaged — near-universal, and not casual

Two-thirds of currently-enrolled students (65%) have used AI for their coursework, and — contrary to the "young person's technology" cliché — adoption is flat across the ages of people actually in school (65 / 66 / 63 / 67% across the native bands 18–20 / 21–30 / 31–40 / 41+, with no age contrast significant under controls); the apparent age gap was really about who is still enrolled (Report 1). The typical student treats it as a tutor more than a ghostwriter — the most common uses are explaining concepts and checking work, with essay-writing ranking second-to-last of eight tasks — and most use it at the margins, though about 18% do "most" or "all" of their schoolwork with it (Report 1). One tool dominates: ChatGPT is the near-universal default, used about weekly, with a long thin tail behind it.

This engagement runs deepest exactly where students have the fewest alternatives. For 17% of student users — 20% of the lowest-income — AI is the help they'd "have no other way" to get; 19% report a learning difference and 69% of those use AI because of it; and among students hesitant to approach their instructors, 68% use AI to ask what they're embarrassed to ask a person (Report 4). AI isn't a gadget bolted onto student life. For a large minority it's load-bearing.

2. Uneasy — and use doesn't fix it

Yet this is not a comfortable embrace. Asked how they feel about the rise of AI, students' top emotions are Concerned (6.4 out of 10), Anxious (5.8), and Curious (5.6); every enthusiastic emotion — Excited (4.7), Optimistic (4.5), Enthusiastic (4.5) — sits at the bottom (Report 7). Students are also more emotionally activated than older adults on nearly every emotion, and the widest gaps are on the apprehensive side (Anxious +1.0, Angry +0.8). Curiosity, not enthusiasm, is what pulls them toward the tool.

The integrative test for this capstone asked whether heavier use dissolves that unease. It does not. As students go from doing none of their schoolwork with AI to doing all of it, excitement rises to roughly two and a half times its starting level (2.75 → 6.71) and curiosity climbs sharply (3.98 → 7.04) — but anxiety barely drops (6.25 → 5.30) (Table 1). Enthusiasm responds several times as strongly to use as anxiety does, and the gap holds under controls: net of party, gender, age, income, and race, each additional step in how much schoolwork a student does with AI predicts a full point more excitement (+1.0 on the 1–10 scale) but only a small drop in anxiety (−0.14) — the enthusiasm association is roughly seven times the anxiety association (Table 2). The heaviest users are the most excited students on campus and, at the same time, still among the most anxious. Because the positive emotions surge while anxiety only drifts, this isn't an artifact of engaged students rating every feeling higher — the pattern is specific: use buys enthusiasm, not peace of mind.

The unease shows up institutionally and socially, too. Students operate in a governance vacuum — only 11% say their instructors mostly encourage AI, 32% say they discourage it, 46% face a mix, and half say it isn't built into any required course — even as the grades that matter most still ride on the written work AI does best (Report 3). And they don't trust each other: students say AI helps them learn (41%) but makes their peers learn less (35%), and picture peers as far more dependent on it than they admit being themselves (44% vs. 18% "most/all"), even while 46% concede their own corner-cutting (Report 2). A generation this engaged is not relaxed — about the technology, their institutions, or each other.

The unease reaches the credential itself, and here the series steps briefly outside the enrolled-student base. Report 5 asks the general adult sample, high-school students, and parents what AI does to the value of a degree, and finds a repricing rather than a collapse: adults lean net-negative on the bachelor's (32% less valuable vs. 23% more) but net-positive on the PhD, and they mark down the humanities (Philosophy −8 net) while marking up engineering and computer science (+17, +15). Crucially, the people making the decision are not walking away — 60% of high schoolers say AI hasn't changed their college plans, and parents lean toward sending their children (27% more likely vs. 12% less), with 58% wanting AI actively taught. The complaint is about content, not enrollment. And it is the enrolled students themselves who register it most sharply: 42% say their program has prepared them poorly for AI or has not addressed it at all, against 14% who feel "very well" prepared. That is the institutional counterpart to everything above — students absorbing a tool on their own, in a governance vacuum, from programs that by their own account are silent on it. It is also where the literacy gap in §3 comes from: nobody is teaching them where the tool breaks.

3. Uncritical — confident, and wrong about the right things

The third leg is the sharpest. Students feel competent — 68% rate themselves skilled with AI, double the rate of older adults — and they correctly grasp what AI is in the abstract (90% know it can be confidently wrong). But on the specific misconceptions that matter when you actually use it for schoolwork, a third to nearly half are wrong: only 63% know an AI citation isn't always real, 67% that it doesn't always know when it's accurate, 56% that it isn't neutral (Report 8). And — the finding that closes the loop with the engagement story — heavier use comes with no better, and by the models somewhat worse, grasp of those failure modes: net of demographics, each step up in how much schoolwork a student does with AI predicts a 7–8 percentage-point lower likelihood of knowing that an AI citation isn't always real, or that AI doesn't always know when it's accurate (Table 2; these are linear-probability coefficients, so they read in percentage points, not odds). As in Report 8, part of that association may reflect a response-style tendency of heavy users to agree with statements, so the robust claim is the absence of any protective effect — experience with the tool does not teach its failure modes. The students leaning hardest on AI are not the most alert to the ways it can betray them; reliance is running ahead of literacy.

That credulity has a self-serving edge. Students are accurate about how many peers use AI (62% estimated vs. 65% actual) but systematically misjudge everything past prevalence — how much peers rely on it, whether it helps them learn — always in the direction that flatters themselves (Report 2). They are confident about their own discernment and wrong about the tool's limits at the same time.

4. The engine underneath: trust, and who's inside it

What organizes all of this is not identity but trust in AI companies — the single strongest predictor of how students feel and behave. It drives excitement up and anxiety down, and it accounts for most of the partisan gap that recurs across the series. The right-leaning end of the spectrum is the AI-embracing end, but the two measures do not peak in the same place: Republicans lead on adoption (79%, against 75% for Strong Republicans and a low of 58% among pure independents), while Strong Republicans lead on intensity (mean 3.05 on the 1–5 none-to-all scale) and on trust (2.82 on the four-point AI-trust scale, against 1.80 among Lean Democrats and ~1.8–2.4 for everyone else). That trust gap, not party as such, is what the partisan pattern is made of — the party coefficients shrink substantially once trust enters the model (Report 6). Trust is low across most of the campus, which is precisely why the embrace is uneasy.

The other axes are secondary but real and consistent. Race is three findings, not one, and they should not be collapsed: net of controls, Black students do a larger share of their schoolwork with AI (intensity +0.34); Black and Hispanic students report more excitement about it (+0.77 for Black students); and Black and Hispanic adults are more likely to see a college degree as holding its value (−0.19 / −0.12 on the degree-devaluation scale, general adult sample). Those are three different outcomes on three different bases, moving in a broadly similar direction — more embrace, more confidence — but each stands or falls on its own. What ties them together negatively is what is absent: there is no racial access gap here, the access shortfall runs along income (Reports 1, 4, 5, 7). The race adoption edge specifically is reference-dependent — significant with White as the implicit reference and not once a White indicator is added — and is read as the absence of a deficit rather than a firm advantage. Gender: men bring more enthusiasm (markedly more excited, heavier use) but not less fear — the gap is far larger on enthusiasm (+1.06 on Excited) than on fear (+0.34 on Scared, women higher), and it is absent on anxiety specifically. Income: shapes access, not attitude. Age, urbanicity, region: essentially inert for behavior — regionally, adoption spans only 59.1% (Northeast) to 67.9% (West), with no region significantly different from another (model F p =.054). The consistent through-line is that the students most comfortable with AI are the ones who most trust it — and that comfort is the exception, not the rule.

Conclusion — what it adds up to, and what to do about it

Put the eight reports together and the portrait is coherent and a little unsettling: students have made AI part of how they learn before they are comfortable with it and before they understand it. They adopt it nearly universally and lean on it hardest where they can least afford alternatives; they feel more concern and anxiety than excitement, and using it more doesn't calm them; and they are confident in their own skill while missing the specific ways the tool fails. The optimistic reading — "digital natives who've embraced AI" — is wrong in both directions: they haven't embraced it so much as absorbed it, and their fluency is thinner than their confidence.

The practical implication is narrow and consistent across the series. More exposure is not the lever — students already have plenty, and exposure brings enthusiasm without either comfort or literacy. The two things that would actually move this generation's relationship with AI are trust — earned, not assumed, and currently low across most of the campus — and targeted literacy about the concrete failure modes (fabricated citations, confident errors, non-neutral output), taught precisely to the confident heavy users who think they already know. A cohort this engaged is not waiting to be sold on AI's capabilities. It is waiting for reasons to believe the technology is on its side, and for someone to teach it where the tool breaks.

Appendix A — Methods

• Population & data. Currently-enrolled students in CHIP50 Wave 38.6 (fielded June 21–July 13, 2026), except where a feeder report uses a broader base (Report 5's degree-value and parent/HS items span the adult, parent, and high-school populations). Survey-weighted; self-reports throughout — this is what students say about their use, feelings, and knowledge, not audited behavior or measured learning.
• What each section draws on. §1 Engaged → Reports 1, 4. §2 Uneasy → Reports 7 (affect), 3 (rules), 2 (peers), 5 (the credential and the preparation gap), plus the new Table 1 analysis. §3 Uncritical → Reports 8 (literacy), 2 (self-serving gap). §4 Trust/axes → Reports 6, and the race/gender/income threads across 1–8. Every figure cited here was independently re-derived and reviewed.
• New analysis (Tables 1–2). Table 1 is descriptive — ai_emotion_2/1/7 (Excited/Curious/Anxious, 1–10) by ai_course_extent (1 = None … 5 = All), currently-enrolled, weighted. Table 2 puts the same relationships through survey-weighted OLS — Excited, Anxious, and the two reliability-knowledge items (ai_quiz_12 citations, ai_quiz_9 accuracy) each regressed on ai_course_extent plus the full demographic set (party7, gender, age band, education, income, urbanicity, race) — confirming the associations survive controls. The emotion-specific divergence (positives up, anxiety down/flat) is the key guard against the response-style acquiescence that confounds the literacy items in Report 8.
• Limits. One wave, cross-sectional — no trend and no measured outcomes (learning, grades); "AI helps me learn" is always a perception. Trust→affect is a decomposition, not a demonstrated causal chain. The knowledge items inherited from Report 8 carry that report's response-style caution: confidence and heavy use are associated with agreeing to declarative statements, which inflates apparent "correctness" on true-keyed items and depresses it on false-keyed ones, so the intensity→literacy coefficients in Table 2 are read as "reliance runs ahead of literacy," not as evidence that heavier users know less. The emotion results are the guard against that confound, since positives and anxiety move in opposite directions. Race coefficients throughout the series are reference-dependent — the direction and significance of a race term shift with the omitted category and the control set, so race differences are reported descriptively and only where they hold across specifications. ---

Appendix B — Data tables

Table 1. Enthusiasm scales with use; unease is sticky (currently-enrolled; weighted mean, 1–10)

Base ≈ 2,190 currently-enrolled students (ai_emotion_2/1/7 × ai_course_extent, 1 = None … 5 = All). Higher = the emotion describes the student more (1–10). Positive emotions rise ~3–4× as much as anxiety falls — and they move in opposite directions, which rules out a "heavy users rate every emotion higher" response-style explanation. Weighted means.

Table 2. What one step more AI use predicts, net of demographics (currently-enrolled; survey-weighted OLS)

Each row is a survey-weighted OLS among currently-enrolled students, regressing the outcome on ai_course_extent (intensity, 1 = None … 5 = All, entered continuously) plus party7, gender, age band, education, income, urbanicity, and race (Black/Hispanic). "Per step" = the change for each one-category increase in how much schoolwork a student does with AI; "across None → All" scales it over the full four-step range. The two emotion outcomes are on the 1–10 scale; the two knowledge outcomes are binary (correct/incorrect), so their coefficients read as percentage points. *** p<.001. The intensity–excitement association (+1.01/step) is ~7× the intensity–anxiety association (−0.14/step), and heavier use predicts lower reliability-item knowledge. These are associations, not causal estimates — excitement and use are mutually reinforcing, and (per Report 8) part of the use → literacy link may reflect response style, so the load-bearing claim there is the absence of any protective effect of experience. SEs model-based.

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.

This capstone collects no new items. Every variable it cites is documented in the Appendix C of the feeder report where it first appears: ai_course, ai_school_use_1–8, ai_course_extent, ai_freq_x1, school_recency (Report 1); ai_learn, ai_learn_others, ai_course_extent2, ai_banned, ai_vs_tutor, ai_study_effort (Report 2); instructor_view, ai_course_integ, ai_assess_matrix_1–6, ai_assess2 (Report 3); ai_equalizer_1–6, ai_ada, ai_ada_use, instructor_comfort, english_first, english_ai (Report 4); ai_deg, ai_degree_fields_1–9, ai_college_forego, parent_view, parent_curric, ai_edu_prep (Report 5); ai_emotion_1–11, pol_trust_ai (Report 7); and ai_proficiency, ai_quiz_1–12 (Report 8).

The two new tables in this report use only items already documented above: ai_emotion_2 (Excited), ai_emotion_1 (Curious), and ai_emotion_7 (Anxious) on the 1–10 scale; ai_quiz_12 and ai_quiz_9 (both keyed False, so "correct" is stored code 2); and ai_course_extent (1 = None … 5 = All) as the intensity predictor.