Generative AI in Higher Education: Inside a 95,000‑Student Study on Cheating and Access Disparities

The Scale and Scope of the Study

The data was collected in spring 2024 through Berkeley’s Student Experience in the Research University (SERU) Consortium — a collaborative network of research universities that pool student survey data to improve higher education outcomes.

The sample size alone makes this study significant. Most prior research on student AI use relied on small, single-institution surveys. At 95,000+ respondents, this is a different category of evidence entirely.

Key baseline numbers from the study:

• ~67% of students reported using generative AI
• ~40% used it monthly or more frequently
• At least 9% of AI users reported using it to cheat

Those numbers sound manageable until you zoom in on the details.

The Cheating Slope Is Real — and Steeper Than Expected

Here’s the finding that should concern every educator and institution: the more frequently students use AI, the more likely they are to cheat with it.

Among students who used AI daily, 26% reported using it to cheat. Among monthly users, that number dropped to just 7%. That’s nearly a fourfold difference based on usage frequency alone.

Chirikov is careful not to claim causation. It’s possible that students already predisposed to cheating simply gravitate toward AI tools more heavily. But the correlation is sharp and consistent — and it points to a structural problem, not just individual bad behavior.

Cheating Is Significant — But Not as Widespread as Feared

One of the study’s more counterintuitive findings is that AI-assisted cheating, while real and growing, is not as rampant as some earlier reports suggested.

Chirikov attributes this partly to timing — the survey captured behavior from early in the AI adoption cycle, before the most capable models became mainstream. He also notes that the study used indirect survey methods to encourage honest reporting, but students still had to self-identify their behavior as non-compliant, which likely produced conservative estimates.

Even so, the numbers are significant enough to demand institutional action. Chirikov’s separate research on grade inflation shows that when students use AI on assignments, entire course grade distributions can shift upward — independent of actual learning.

That’s a credentialing problem. And it’s already happening.

STEM vs. Non-STEM: Discipline Matters

The study found meaningful differences in AI misuse across academic disciplines. Non-STEM students were more likely to cheat with AI than their STEM counterparts.

This makes intuitive sense. Writing-heavy assignments — essays, literature reviews, reflective papers — are more easily outsourced to a language model than a calculus problem set or a lab report. But it also signals that blanket university-wide policies are the wrong tool for this problem.

Chirikov and his co-authors argue that AI governance needs to be discipline-specific, and in many cases course-specific. What responsible AI use looks like in a creative writing seminar is fundamentally different from what it looks like in a software engineering course or a chemistry lab.

The Assessment Redesign Challenge

One response gaining traction is moving all assessments into controlled environments — proctored oral exams, handwritten in-class tests, live coding sessions. These formats are harder to fake with AI.

But Chirikov flags a real cost here. Controlled, time-limited assessments only capture a narrow slice of what universities actually teach. Research skills, sustained critical thinking, iterative writing, and complex problem-solving all require prolonged engagement with material — the kind of struggling intellectually that is itself part of the learning process.

If institutions over-correct toward AI-proof assessments, they may inadvertently strip out the most valuable parts of a university education.

The Access Gap: A Bigger Problem Than Cheating

If the cheating findings are concerning, the access disparity findings are alarming.

The study found that low-income, racially underrepresented, and female students were significantly less likely to use generative AI than their peers. This isn’t a minor statistical blip — it’s a structural gap with long-term career implications.

Here’s why it matters so much: employers are increasingly looking for graduates with hands-on AI experience. Students who graduate without meaningful AI proficiency may find themselves at a disadvantage in the job market — not because of their skills or intelligence, but because they lacked access to the AI tools during their formative years.

What Universities Should Actually Do

The study doesn’t just diagnose problems — it points toward solutions, even if those solutions are hard.

1. Abandon blanket AI bans. They don’t work, and they may actively harm students who need AI proficiency for their careers. Students will use AI regardless. The question is whether institutions help them use it responsibly.

2. Develop discipline-specific policies. What counts as appropriate AI use in a nursing program is different from what it means in a computer science course. Policies need to reflect that reality.

3. Redesign assessments thoughtfully. The goal isn’t to make AI impossible to use — it’s to design assessments that actually measure what students know and can do. That requires investment, creativity, and faculty support.

4. Address the access gap directly. Universities should consider subsidizing access to AI tools for low-income students, the same way they subsidize textbooks, software licenses, and computing resources.

5. Teach AI literacy as a core competency. Not just how to use tools, but how to evaluate their outputs, understand their limitations, and make informed decisions about when to use them.

What This Means for Students Right Now

Chirikov offers a simple but powerful self-check for students using AI on coursework:

“Could I explain this without the tool? Could I do a similar task on my own tomorrow? Did AI help me understand the material better, or did it mainly help me finish faster?”

These questions cut to the heart of the issue. AI can produce a polished essay. It cannot build the underlying skill that essay was meant to develop. Students who outsource their learning to AI may graduate with credentials that don’t reflect their actual capabilities — and that gap will surface in the workplace.

The experimental evidence Chirikov cites is sobering: people learn significantly worse with AI assistance and fail to develop durable skills compared to learning without it. The short-term grade boost comes at a long-term cost.

The Bigger Picture: Trust in Higher Education

There was already a trust deficit in higher education before generative AI arrived. Concerns about credential inflation, rising costs, and questionable return on investment were well-established.

AI adds a new dimension to that crisis. If grades are increasingly AI-inflated, and if institutions can’t reliably assess what students actually know, the value of a degree becomes harder to defend.

How universities respond to this challenge over the next few years will shape public trust in higher education for a generation. The evidence from this study is clear: the response needs more resources, higher prioritization, and a willingness to rethink assessment from the ground up.

The Takeaway

This study is the most rigorous look we have at how generative AI is reshaping undergraduate education — and it surfaces two distinct crises running in parallel.

The first is academic integrity: a real, measurable cheating problem that scales with AI usage frequency and resists simple policy fixes.

The second is equity: a growing divide between students who can afford powerful AI tools and those who can’t — with direct implications for career readiness and economic mobility.

Neither problem has an easy solution. But both demand that universities, educators, and AI tool builders take them seriously now, before the next generation of models makes both crises significantly harder to manage.

The data is in. The question is what institutions do with it.