AI in Higher Education: Cheating, Equity Gaps Behind the Hype

The adoption story, and what it leaves out

By most headline measures, generative AI has won the education sector. IDC research cited in Microsoft's 2025 AI in Education Report puts adoption among education organizations at 86%, which it calls the highest of any industry 5. A Coursera report from February 2026 found that four in five students say AI has improved their academic performance 5. A peer-reviewed randomized controlled trial published in Scientific Reports in 2025 found that an AI tutor outperformed traditional classroom instruction by a wide margin 5.

Those numbers are real, but they describe uptake and perceived benefit. They say little about whether the work students hand in still reflects what they know. They also say little about who benefits and who pays a price. Two recent lines of research, one on campus integrity and one on college admissions, fill in that gap.

A validity problem, not just a cheating problem

A May 2026 study published in Science and publicized by Cornell and UC Berkeley drew on survey responses from more than 95,000 students at 20 public research universities 4. A summary of what appears to be the same research describes nearly 100,000 US university students. It reports that roughly a quarter of daily generative AI users admit to academic cheating 2. The OECD's AI incident monitor logged the findings as a realized harm rather than a hypothetical risk 2.

The researchers' central argument goes beyond the cheating rate. They say widespread AI use and misuse undermines the validity of assessment. Faculty can no longer be confident that a grade reflects what a student can actually do 4. That framing matters. A cheating problem invites enforcement. A validity problem means some types of assignments may simply have stopped measuring learning.

Faculty experience supports the concern. One national survey found 73% of instructors say they have personally handled AI-related integrity cases 3. Professors at Brown and Alcorn State drew wide attention after traps they set suggested most of their students had used AI on major assessments 3. Autonomous AI agents have also shown they can complete entire online courses, and earn top marks, without the enrolled student doing anything 3. A January 2026 survey by Elon University and the American Association of Colleges and Universities found widespread faculty worry about overreliance, weaker critical thinking, and integrity lapses 4.

Detectors out, redesign in

The obvious technical fix has largely failed. A growing number of universities have banned AI detection tools as unreliable. Faculty are left facing rampant misuse with fewer ways to prove it 3. Institutions are instead pushing professors to rework assessments so AI cannot easily complete them 3. Commentators increasingly describe redesign, not policing, as the real solution 4.

The equity gap hiding in plain sight

The second line of research is a separate 2026 study from Cornell and Carnegie Mellon. It analyzed more than 81,000 college applications 1. Lower-income applicants were 28% more likely to lean heavily on AI. When they did, their admission odds fell by 83%, compared with a 62% drop for wealthier applicants using AI the same way. Those results held after controlling for GPA and test scores 1.

The proposed explanation is that the gap comes less from how much AI students use than from whether an expert reviews the AI's output 1. Wealthier applicants are more likely to have counselors, tutors, or parents who can spot generic or flawed AI prose. Lower-income students often lack that check. This shifts the inequality debate away from subscriptions, devices, and broadband, the usual focus, toward access to human judgment 1.

Reading the evidence together

The two studies measure different things, and they should not be merged into a single finding. The Science work concerns integrity on campus. The Cornell–Carnegie Mellon work concerns admissions outcomes. Still, they point in the same direction. AI's effects depend heavily on the system around it. The randomized tutoring trial and the student satisfaction surveys show what AI can do in well-designed settings 5. The integrity and admissions data show what happens when AI is dropped into systems built for a pre-AI world.

My reading is that the adoption statistics are the least informative part of this story. Near-universal uptake is now a given. The open questions are whether institutions can rebuild assessments that measure learning again, and whether expert guidance on AI use becomes a new marker of privilege. If colleges redesign coursework but leave the guidance gap alone, they risk fixing one problem while deepening the other. Treating AI literacy and supervised use as a support service, rather than a luxury, looks like the logical next step. That is an inference from the evidence, not something the studies themselves prescribe.