The Honor-Roll Offloaders
The students leaning most heavily on GenAI aren't the ones failing. They're the ones with the highest GPAs—and that should worry us more.
Picture the student you’d least suspect of cheating. Not the one cutting corners to scrape a C, but the one with the immaculate transcript, the one professors write recommendations for without hesitation. Now imagine her ChatGPT history.
A 2026 preprint, “Patterns of Student Cognitive Offloading to AI in Higher Education,” analyzed 3,047 ChatGPT messages from 46 undergraduates and found that the heaviest use was not among the strugglers but among the high performers. The honor-roll students were offloading the most. (It’s a preprint, not yet peer-reviewed, and the sample is small — but the pattern is worth sitting with.)
This breaks the story we’ve been telling ourselves. The campus AI panic has always cast the cheater as someone desperate: overwhelmed, underprepared, gaming a system they can’t otherwise beat. That student exists. But the data point somewhere more uncomfortable: for the best students, AI dependence is rarely a desperate move and more often an optimization strategy.
And why wouldn’t it be? We built the incentives ourselves. We told a generation that the grade is the goal, that the transcript is the prize, that the point of the assignment is the artifact you hand in. A bright, rational student hears that message and does the obvious thing: she finds the fastest route to the artifact. AI is simply the fastest route yet. She understands the game perfectly.
There’s a particular cruelty in this for the conscientious student. She isn’t lazy; she’s overloaded — five courses, a job, the ambient dread that any slip will cost her the future she’s been told depends on this transcript. AI doesn’t tempt her the way it tempts the struggling student; it rescues her. It’s the only thing offering to give back the hours hyperachievement stole, and it asks for nothing except the part of the work that was invisible anyway. Who refuses that bargain? Only someone who was told what it actually costs, and almost certainly, no one has told her.
And the cost is deferred, which is what makes it so easy to pay. The thinking she skipped this week shows no visible deficit. The essay is fine. The grade is fine. The deficit surfaces years later, diffuse and unattributable: an adult who can generate a competent document and cannot, when it matters, tell a sound argument from a slick one. The bill arrives with no name on it.
Now look at the same problem from the registrar’s side, where it stops being anecdote and becomes a ledger. This spring a University of California, Berkeley team analyzed 500,000 grades across eight years and found that in writing- and coding-heavy courses, the share of A’s rose about 30 percent after ChatGPT — and the jump lived almost entirely in take-home work, not in proctored exams. As the study’s author put it, you now have “a C student who is now an A student.” The grade went up; the knowing didn’t. EDUCAUSE Review named the paradox in 2025: “better results, worse thinking.”
So here’s my claim, stated plainly. The honor-roll offloader is not a discipline problem. She is a measurement problem. We spent a century building an education system that grades outputs, because outputs are easy to grade and were reasonably hard to game. AI just revealed that a graded output is no longer evidence of anything happening inside a head, and now the ledger says so at scale. Our best students were the first to notice. Of course they were; noticing is what made them our best students.
The institutions have noticed too, and their alarm is the tell. Princeton is ending 133 years of unproctored exams; Harvard’s faculty voted to cap A’s at a fifth of each class. Both are confessions that the transcript has stopped certifying what it claims to. But capping grades treats the symptom. The real question is the one the Berkeley author asks: do grades still tell us what students can actually do?
What we should be developing is reasoned judgment: the capacity to sit with an argument’s claim and reasoning and decide whether it holds. You cannot offload that to a machine, because the machine was trained on the internet, where reasoned judgment was never required to post. Outsource your thinking to ChatGPT, and you become more reliant; learn to interrogate what it hands you, and you become harder to fool. Same tool, opposite results.
The fix isn’t detection; that’s a tax on trust the cleverest students will always outrun. It’s assignments where overreliance carries a penalty, and the thinking is the deliverable: done in front of you, explained out loud, and graded explicitly, as the thing itself. If we make the thinking the deliverable, the shortcut stops being short.
Until we do that, the most capable people in the room will keep making the most rational possible choice. We won’t like where it leads them, and for now, the transcript won’t even tell us it happened.
A question for the comments: If your best students are the heaviest AI users—and the grade can no longer reveal who learned what—is that a cheating problem or a curriculum problem, and what would you actually change?
Read deeply,
Dr. Genevieve