The Baseline: Who’s Using AI and How Often
About two-thirds of surveyed students reported using generative AI tools. Nearly 40% used them monthly or more frequently.
That’s a significant adoption rate for a technology that only entered mainstream consciousness in late 2022. By spring 2024 — when this data was collected — ChatGPT and similar tools had already become embedded in student workflows across disciplines.
But usage frequency isn’t uniform, and that variation matters enormously for understanding what comes next.
The Cheating Gradient: More Use, More Risk
Here’s the finding that should make every university administrator pay attention.
Among daily AI users, 26% reported using AI to cheat. Among monthly users, that number dropped to just 7%. The correlation is stark: the more frequently a student uses AI, the more likely they are to cross the line into academic dishonesty.
Chirikov is careful not to claim causation. It’s possible that students already prone to cheating are simply drawn to AI tools more readily. But the trend is consistent and worrying regardless of direction.
What’s driving it? A perfect storm, according to Chirikov. Selective research universities create intense grade pressure — GPAs determine internships, graduate school admissions, and career trajectories. Meanwhile, AI tools have made it trivially easy to generate a polished assignment in under 30 minutes instead of pulling an all-nighter. When those two forces collide, academic integrity becomes collateral damage.
Non-STEM Students Cheat More With AI
The study also found meaningful differences across disciplines. Non-STEM students were more likely to report using AI to cheat than their STEM counterparts.
This makes intuitive sense. Writing-heavy assignments in humanities and social sciences are easier to outsource to a language model than a calculus problem set or a lab report. But it also signals that blanket university-wide AI policies will miss the mark — what responsible AI use looks like in a literature course is fundamentally different from what it looks like in a computer science program.
Why Banning AI Isn’t the Answer
The instinct to ban AI across the board is understandable. It’s also counterproductive.
Students will continue using AI regardless of policy. Many already use it to learn — asking questions they’d be too embarrassed to raise in class, getting explanations of difficult concepts on demand, working through problems iteratively. Cutting off that access doesn’t eliminate the behavior; it just drives it underground.
There’s also a career readiness argument. Employers increasingly expect graduates to be AI-proficient. A blanket ban during college doesn’t prepare students for that reality — it leaves them behind.
The detection arms race reinforces this point. AI detection software is improving, but so are “AI humanizers” — tools designed to make AI-generated text look human-written. Chirikov describes it plainly: a cat-and-mouse game with no clear winner. Institutions that bet their integrity strategy on detection alone are building on sand.
What Actually Works: Discipline-Specific Policies
The study’s core recommendation is that academic departments — not central administrations — need to own their AI policies.
Some programs are moving toward proctored oral exams and handwritten in-class assessments. These work for testing a narrow band of skills in a controlled timeframe. But research universities exist to develop something broader: the ability to engage deeply with material over time, to struggle intellectually, to iterate.
If assessment design shrinks to only what can be tested in a 90-minute proctored session, universities risk losing the very thing that makes them valuable.
The harder, more important work is redesigning assessments that require genuine demonstration of understanding — outputs that AI can’t convincingly fake because they demand personal synthesis, lived experience, or real-time reasoning. That’s not easy. But it’s the right direction.
The Access Gap Nobody Is Talking About Enough
Chirikov calls this the most important finding in the study — more important, even, than the cheating data.
Low-income students, racially underrepresented students, and female students are all less likely to use generative AI than their peers. This isn’t a minor statistical blip. It’s a structural inequality baking itself into the next generation of the workforce.
Here’s the mechanism: premium AI tools — the ones with stronger capabilities, fewer usage limits, and more reliable outputs — cost money. Students from wealthier families can afford subscriptions to advanced models. Students without those resources are left with free-tier tools that are slower, less capable, and more restricted.
The result is an AI proficiency gap that tracks directly onto existing socioeconomic and racial divides. And as AI becomes more central to hiring decisions, that gap translates into a career disadvantage that has nothing to do with actual skill or intelligence.
The Grade Inflation Complication
There’s another layer here that compounds the problem. Chirikov’s related research on grade inflation shows that when students use AI on assignments, course grades can become inflated relative to actual knowledge.
If wealthier students are using more powerful AI tools and getting better grades as a result, the credential itself becomes less meaningful — and the students who couldn’t access those tools are doubly penalized. They get lower grades and a devalued degree.
What This Means for AI Tool Adoption in Education
From a tools ecosystem perspective, this study surfaces a few clear signals.
Adoption is real and accelerating. Two-thirds of undergrads at major research universities were already using generative AI in spring 2024. That number is almost certainly higher now.
The free-versus-paid divide is consequential. The quality gap between free and premium AI tools isn’t just a product feature — it’s becoming an equity issue with measurable downstream effects on education and employment.
Detection tools have limited utility. Institutions investing heavily in AI detection software are likely to find diminishing returns as humanization tools keep pace. The smarter investment is in assessment redesign.
Discipline-specific tools and policies will outperform generic ones. The variation in how students use AI across subjects suggests that the most effective tools — and the most effective policies — will be those built with specific use cases in mind.
The Questions Every Student Should Be Asking
Chirikov closes with a practical framework for students navigating this landscape. It’s worth repeating directly.
Before submitting AI-assisted work, ask yourself:
- 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 aren’t rhetorical questions. They’re diagnostic. The difference between AI as a learning accelerator and AI as a skill-atrophying shortcut often comes down to whether students are asking them at all.
The Bigger Picture
Higher education was already facing a trust deficit before generative AI arrived. This study adds urgency to a problem that institutions can’t afford to defer.
If grades no longer reliably signal competence — because AI can inflate them — and if access to AI tools tracks socioeconomic status, then universities are at risk of producing credentials that mean less while simultaneously widening the gaps they’re supposed to close.
The data from 95,000 students is clear: this isn’t a future problem. It’s a present one. And the institutions that treat it as such — investing in redesigned assessments, discipline-specific policies, and equitable AI access — will be the ones that emerge from this transition with their credibility intact.
The rest will be playing catch-up in a game that’s already well underway.
Comments (0) No comments yet
Want to join this discussion? Login or Register.
No comments yet. Be the first to share your thoughts!