A Leveler With an Asterisk: AI, Access, and Equity in School

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 "what would you do without AI" battery (ai_equalizer) and the disability-use item (ai_ada_use) were asked only of currently-enrolled AI users (n ≈ 1,433 and ~259); the disability, English-language, and adoption items were asked more broadly (n ≈ 2,200); the judgment-free-practice item (instructor_comfort_ai) was asked only of AI users who are less than "very comfortable" approaching instructors (n ≈ 787).

All percentages are survey-weighted to national demographic targets; unweighted N is a reliability cue only. Figures are self-reported. Income is grouped into three bands (lower / middle / upper). Because two of the equity items rest on small filtered bases, they are reported as toplines with income cuts only, not finer breakdowns.

Key Takeaways

Introduction

The hopeful case for AI in education is that it democratizes help. A private tutor costs money; office hours require confidence and good timing; a good writing coach is a luxury. A chatbot, the argument goes, gives every student a patient, always-available, judgment-free version of all three — leveling a playing field that has always tilted toward students with money, connections, and confidence. The skeptical case is the mirror image: that the students who already have the most will simply use AI to pull further ahead, and that access to the good tools will track the same lines as access to everything else.

CHIP50's Wave 38.6 supplement lets us weigh both. It asks students what they would do without AI, whether they lean on it because of a disability, whether they use it to ask questions they would not ask a professor, and whether they use it to write in English — and it can set those answers against who is and isn't using AI in the first place. The picture that emerges supports neither the pure-equalizer nor the pure-inequality story. For the students who use AI and lack alternatives, it genuinely functions as access. But it is reaching those students least — a limitation a major new study of 95,000 undergraduates calls, if anything, "even more important than the cheating part."¹

The Irreplaceable Option

For a meaningful minority of students, AI is not one source of help among many — it is the only one they can reach. Asked what they would do if they lost access to AI, 17% of currently-enrolled student AI users say they would "have no other way to get that level of help." That share is highest among lower-income students (20%), versus 15% of middle-income and 17% of upper-income students (Table 1).

The texture of the alternatives sharpens the point. The fallbacks that cost money are named disproportionately by students who have money: hiring a private tutor rises from 21% among lower-income students to 26% among wealthier ones, and paying for a study service like Chegg or Course Hero climbs steadily from 14% (lower-income) to 24% (upper-income). Free online resources — Google, YouTube, Wikipedia — are the near-universal backstop (about 69% across the board). In other words, when wealthier students imagine losing AI, more of them can fall back on paid human help; when lower-income students imagine it, more of them are left with nothing beyond a search engine. AI is the free tutor standing in for the paid tutor they cannot afford.

Comparative context. This is the mechanism the accommodation and access literature has in mind when it calls AI "assistive": for students without paid alternatives, it substitutes for help that used to require money or social capital.¹

But Access Runs Uphill

The catch is that the students who would gain most from a free tutor are the least likely to be using one. Adoption of AI for schoolwork rises steadily with income: 60% of lower-income currently-enrolled students have used it, versus 66% of middle-income and 77% of upper-income students (Table 2) — a 17-point gap. The have-nots, for whom AI is most nearly irreplaceable, are also the least likely to have picked it up.

This is not a CHIP50 artifact. The largest study to date of undergraduate AI use — 95,000-plus students across 20 research universities, published in Science in 2026 — finds the same pattern and worse: low-income, racially underrepresented, and female students use AI less, and wealthier students "can access advanced AI tools with stronger capabilities and fewer usage limits" while others "may only be able to use free AI tools."¹

One dimension of that divide looks different in these data than in the national study, though: race. The Science study reports that racially underrepresented students use AI less; CHIP50 does not reproduce a racial access gap. Raw adoption among currently-enrolled students is about the same for Black (64%) and Hispanic (64%) students as for White students, and higher for Asian students (72%). And because Black students are disproportionately lower-income — the very thing that depresses adoption — holding income constant, Black and Asian students are at least as likely to use AI as otherwise-similar White students, and by some model specifications more likely (that "more" is reference-dependent, so it is best read as the absence of a racial access deficit rather than a firm advantage). The report's central reliance measure points the same way: the share who say they would have "no other way" to get help shows no racial difference once income is in the model — it is modestly higher among lower-income users (20% vs. 15%) as a descriptive gradient, though income itself does not reach significance on that particular item. In this sample the access shortfall runs along income, not race — Black and Hispanic students adopt AI at about the same rate as White students and Asian students at a higher rate, and all three groups do a larger share of their schoolwork with it (a pattern the flagship and affective reports develop). The Science study's warning still bites on tool quality — who can pay for the more capable models — but on simple access, the racial gap it documents does not appear here.

Notably, the access gap is also not geographic: adoption is essentially flat across urban, suburban, and rural students (63–66%) and across Census regions, where the whole range runs from 59.1% (Northeast) to 67.9% (West) with no region significantly different from another. So the students on the wrong side of the line are defined by what they can pay, not by where they live. The leveling is real where it happens, but it happens least where it is needed most, and the quality gap — better models for those who can pay — threatens to widen as premium tools improve. AI's equalizing potential, on this evidence, is genuine but only half-realized.

An Accommodation

For students with a learning difference, AI's access function is most literal. Among currently-enrolled students, 19% report a disability or difference that affects their learning (another ~5% prefer not to say). And among those who both report such a difference and use AI for schoolwork, 69% say they use it because of that difference — a direct statement of AI functioning as assistive technology (Table 3). This echoes, on the cleaner currently-enrolled base, the central finding of this project's companion accommodation report, and it lands in the middle of an active legal debate: the U.S. Office for Civil Rights has begun asking whether denying students with disabilities access to AI could itself raise questions under the Americans with Disabilities Act.²

Here the equity logic runs the opposite way from income: rather than a gap to be closed, this is a group for whom AI is already doing accommodation work — quietly, and largely outside the formal disability-services process.

Figure 3. AI as accommodation — two stat callouts: 19% report a learning difference, and 69% of those who use AI use it because of that difference. Base: currently-enrolled students, unweighted n ≈ 2,286 (ai_ada); currently-enrolled AI users with a learning difference, unweighted n ≈ 259 (ai_ada_use). CHIP50 Wave 38.6, fielded June 21–July 13, 2026. Weighted estimates.

Judgment-Free Practice — and a Non-Finding on Language

AI also serves as a low-stakes place to ask the questions students are embarrassed to ask a person. The item was routed narrowly: it was put only to currently-enrolled AI users who describe themselves as less than "very comfortable" approaching their instructors (n ≈ 787). Among them, 68% say they use AI to ask questions they would feel uncomfortable asking their instructor.

Getting the scope right matters, because the filter is less restrictive than it sounds. Only 44% of currently-enrolled students say they are "very comfortable" approaching an instructor; the other 56% — a majority — are somewhat comfortable or less, and it is the AI users within that majority who were asked (Table 4). At the same time, outright discomfort is rare: 86% are at least "somewhat" comfortable, and only about 14% say "not very" or "not at all." So this is not a student body hiding from its professors. It is a large middle of students who are fine with their instructors but not fully at ease with them, and who now have somewhere else to take the question they would rather not ask out loud.

One equity dimension we expected to find is not in the data. AI help with English-language writing is common — 39% of students use AI to create or edit English-language text for schoolwork — but it is not concentrated among non-native English speakers, who use it at essentially the same rate as native speakers (38% vs. 39%). Whatever AI is doing on the writing front, on this measure it is a general-purpose writing aid, not specifically a language equalizer for the roughly 7% of currently-enrolled students who say English is not their first language. We report the null because the leveler story is only credible if its limits are stated too.

Conclusion

Whether AI is an equalizer depends entirely on where you stand when you ask. For the student who cannot afford a tutor, or who learns differently, or who is too anxious to raise a hand, AI is often exactly what its optimists promise: patient, free, and available — the help that used to require money, a diagnosis, or nerve. Seventeen percent of student users say they would have no other way to get help at that level, and they are disproportionately the students with the least. That is a real leveling function, and it should not be dismissed.

But leveling requires that the tool reach the people it could level for, and here it falls short. Adoption still rises with income, the better tools cost money, and the largest study of the question finds the access gap widening rather than closing. AI is a leveler with an asterisk: a genuine equalizer for those who use it, reaching least those who need it most. The policy task it implies is not to celebrate the equalizer or debunk it, but to close the gap between its promise and its reach — through access to capable tools, not just any tools, for the students currently on the wrong side of the line.

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_equalizer_1–6 and ai_ada_use: currently-enrolled AI users (n ≈ 1,433 and ~259). ai_course (adoption), ai_ada, english_ai/english_first: currently-enrolled (n ≈ 2,200). instructor_comfort_ai: currently-enrolled AI users who are less than "very comfortable" approaching instructors (n ≈ 787); instructor_comfort itself, which defines that filter, was asked of all currently-enrolled students (n ≈ 2,288).

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

Measures (verified coding). ai_equalizer_1–6 — "If you had no access to AI tools, what would you do?" (select-all; 1 = selected); options include "no other way to get that level of help," hire a private tutor, ask family, free online resources, pay for a study service, office hours/study groups. ai_course (adoption) = % Yes. ai_ada — "Do you have a disability or difference that affects your learning?" (1 = Yes, 2 = No, 3 = Prefer not to say). ai_ada_use — among learning-different AI users, "Do you use AI to learn or complete schoolwork because of this?" (1 = Yes). instructor_comfort_ai — among students less than "very comfortable" with instructors, "Do you use AI to get help with questions you'd feel uncomfortable asking your instructor?" (1 = Yes). english_ai — "Do you use AI to create/edit text in English for schoolwork?" (1 = Yes); english_first — "Is English your first language?" (1 = Yes, 2 = No). Income banded from income_cat_10.

Appendix B — Data tables

Table 1. "If you had no access to AI, what would you do?" by income (currently-enrolled AI users; weighted % selecting)

Base ≈ 1,433 currently-enrolled AI users. Select-all; columns do not sum to 100%. Weighted %.

Table 2. Ever used AI for schoolwork, by income (currently-enrolled; weighted %)

Base ≈ 2,214 currently-enrolled students. Weighted %. Adoption rises with income — the access gap.

Table 3. AI as accommodation (currently-enrolled; weighted %)

Bases: ai_ada n ≈ 2,286; ai_ada_use n ≈ 259. Weighted %. Compare the companion accommodation report (broad student block: 14.7% / 74.2%).

Table 4. Comfort approaching an instructor (currently-enrolled; weighted %)

Base ≈ 2,288 currently-enrolled students (instructor_comfort; 1 = very … 4 = not at all). Less than "very comfortable" = 56%; at least "somewhat" = 86%. The instructor_comfort_ai follow-up (68% yes) was asked only of the AI users within the 56%, n ≈ 787. Rows may not sum to exactly 100 due to rounding. Wave 38.6.

Demographic robustness (SES, urbanicity, region)

SES is the report's spine (adoption rises with income, 60→77%). The race findings — that the access gap is income rather than race, and that racial-minority students are no less AI-embracing net of income — are integrated in the body ("But Access Runs Uphill"); the "more" is a reference-dependent adoption suppression effect (Black OR 1.26 with White as the implicit reference, non-significant once race_white is added), so it is read as the absence of a racial access deficit rather than a firm advantage. On the "no other way" reliance item, race makes no difference (all race terms non-significant); income does not reach significance either, so it is a descriptive lower-income tilt, not an "income-driven" effect. Urbanicity and Census region are null — regional adoption spans only 59.1% (Northeast) to 67.9% (West), no pairwise difference significant (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 student base.
• ai_equalizer_1–6 (select all) — "If you had no access to AI, what would you do?" (1 = Selected). Options: 1 no other way to get that level of help · 2 hire a private tutor · 3 ask a knowledgeable family member · 4 free online resources · 5 pay for a study service (Chegg/Course Hero) · 6 unpaid office hours or study groups. Select-all: shares do not sum to 100%.
• ai_course — "Have you ever used AI for schoolwork?" (1 = Yes; 2 = No; 3 = Not sure).
• ai_ada — "Do you have a disability or difference that affects your learning?" (1 = Yes; 2 = No; 3 = Prefer not to say).
• ai_ada_use (asked of those answering Yes) — "Do you use AI to learn because of it?" (1 = Yes; 2 = No; 3 = Prefer not to say).
• instructor_comfort — "How comfortable are you asking your instructors questions?" (1 = Very; 2 = Somewhat; 3 = Not very; 4 = Not at all comfortable).
• instructor_comfort_ai (asked of those less than "very comfortable") — "Do you use AI for questions you'd be uncomfortable asking an instructor?" (1 = Yes; 2 = No).
• english_first — "Is English your first language?" (1 = Yes; 2 = No).
• english_ai — "Do you use AI to create or edit English text for schoolwork?" (1 = Yes; 2 = No).
• income_cat_10 — derived 10-category household-income banding; collapsed to lower / middle / upper for presentation.