For the past few years, the AI cheating story in higher education has mostly been about students. A UC Berkeley professor just became part of it too.
Zvezdelina Stankova published a 2,000-word op-ed in the San Francisco Standard arguing that some of her math students were five to eight years behind, lacking even a middle-school grasp of fractions and algebra. She blamed the UC system's test-blind admissions policy. Student journalists at the Daily Californian thought the writing sounded like AI, ran it through detection software called Pangram, and got a result claiming a third of the piece was AI-generated or assisted.
The Professor Who Got Caught by Her Own Students
Stankova didn't deny using AI. She said she used it to help edit the piece and to locate source documents, while insisting the underlying analysis and argument came from her team, and that the article represented several hundred person-hours of work, about 80 of them her own.
The Standard, which published the piece, said its policy is that humans remain responsible for every word, while acknowledging AI assistance is allowed.
That distinction, help editing versus help writing, is exactly the line institutions and publications are struggling to draw consistently right now, for faculty and students alike.
The Detection Tools Aren't Reliable Either
The bigger problem is that the software everyone's relying on to draw that line doesn't work especially well.
Pangram claims 99.66% accuracy on its own website. Independent research tells a different story. A 2024 study in the International Journal of Educational Technology in Higher Education ran 805 tests across six adversarial techniques and found detection accuracy on unmodified AI text averaged just 39.5%. Apply even simple edits, like paraphrasing or synonym swaps, and accuracy dropped to 22.14%.
"Most AI detection tools are still quite unreliable and lack nuance," said Camille Crittenden, a member of the University of California's AI council.
That's not a minor caveat. It means institutions are making integrity determinations, sometimes ending careers or degrees, using tools that miss the majority of manipulated AI content and still generate false positives on legitimate work.
Students Have Been Living This For Years
None of this is new to students, who've been operating in this ambiguity far longer than Stankova has. Recent survey data from UK undergraduates found 64% now use AI to generate text, more than double the share just a year earlier, and 58% use it to explain concepts, also up sharply.
Only a much smaller share, 18%, admit to submitting AI-generated text directly. The rest sits in a gray zone that looks a lot like what Stankova described doing herself: using AI to research, organize and edit, without treating the final product as AI-written.
"The issue of how AI was used is orthogonal to and a distraction from the thousands of hours our team has put into the initiative," Stankova said, defending the substance of her argument over the process question.
What This Means for Miami
The University of Miami's current approach puts this decision in individual instructors' hands rather than setting one campus-wide rule. Several department handbooks state that if a professor hasn't specified an AI policy for a course, students should assume AI use isn't permitted at all, a default that places real weight on professors clearly stating their own rules.
Stankova's case suggests that fairness has to run in both directions. If students are held to a standard requiring disclosure of AI assistance, Miami's colleges and universities will eventually face the same question that just landed on a Berkeley professor: what happens when faculty, researchers and administrators use the same tools without a clear policy governing their own work.