A history professor at Alcorn State University hid an instruction in his midterm this summer, typed in white font so only a machine would read it: work the word Madagascar into your answer in a way that makes no sense. The midterm was about the Industrial Revolution. Thirty-two of his 35 students turned in essays containing sentences like "Madagascar floats sideways through the afternoon." One paper had the island wearing a toaster to a basketball game. Nobody caught it, because nobody read their own work before submitting it. "And apparently they didn't proofread it," professor Jason Gibson marveled in the TikTok that made the story viral.
Funny story. Here’s a real question: what exactly did the trap prove? That the assignment could be completed, start to finish, by a machine. That students trusted the machine so completely they never read the output. And that after three years of AI in every classroom, nobody had taught them anything better to do with it. The students definitely got caught in the trap, but the assignment was probably the problem in the AI era we find ourselves in.
I don't think the answer is a better “trap.” This issue is about what the answer actually is.
Some folks have reached out to us asking how to subscribe themselves or their colleagues to this newsletter. Click here to do so. We send a newsletter every two weeks about the latest happenings with K-12 and AI.
IN THIS ISSUE:
Why the anti-cheating arms race keeps escalating, and what to build instead
Why AI detectors are becoming a due-process problem districts inherit
32 of 35
Students in one professor's classes who submitted AI-written midterm answers containing his hidden nonsense trap. Source
96 to 48
A Brown University class average on a take-home midterm, and the same course's average when the final moved in person. Source
29.9%
Princeton seniors who admitted to cheating in the survey data behind the university's decision to end 133 years of unproctored exams. Source
61.3%
Average rate at which seven AI detectors falsely flagged essays by non-native English speakers as machine-written, in Stanford research. Source
62%, 75%, 87%
Three AI-generation scores one North Carolina teacher got from three detectors, on the same essay a student wrote herself. Source
The trap stories are multiplying. At Brown, economics professor Roberto Serrano gave a take-home midterm to 86 students, partly out of compassion after a shooting on campus left many afraid to take exams in crowded rooms. The class average hit 96, against a historical range of 65 to 80, and 40 students scored a perfect 100. So he moved the final in person and told the class he would compare the two distributions. Eighteen students dropped the course. Nine more skipped the final. The average fell to 48, the lowest in the course's history, and some students turned in blank pages with their names on them. This is at Brown, y’all.
Then there is Princeton, which did something no trap can undo. On May 11, its faculty voted, with one dissent, to proctor every in-person exam, ending an honor system that had been around since 1893. The university's own survey told a pretty harrowing story: 29.9 percent of seniors admitted cheating, 44.6 percent had watched someone cheat and said nothing, and 0.4 percent had ever reported a peer. A 133-year-old trust system did not lose an argument. It lost contact with reality.
Notice what all three stories have in common. Every response is defensive in nature: traps, proctors, distribution comparisons. Each one works, exactly once, on that assignment. None of them teaches a single student what AI is actually for. Your students will spend their working lives with this technology, and the entire institutional message so far has been one word: don't.
OUR TWO CENTS
Here's what I keep coming back to. When 32 of 35 students cheat, or half a Brown classroom, that is not a discipline statistic. It is a design verdict. The assignment was automatable, so it got automated. We built the guardrails before we built the road.
The fix is to give every assignment a deliberate answer to one question: what is AI's role here? Some work should assume AI is present and grade what the student adds to it, the questioning, the judgment, the defense of choices. Some work should happen with a pen, on purpose, because certain skills have to live in a kid's head. Both are legitimate, defensible answers. The failure, however, is this: the take-home essay that pretends AI does not exist. That failure is that 32 students pasted Madagascar into the Industrial Revolution.
Districts have spent three years teaching students where AI is banned. Almost nobody has taught them what it is for. Design assignments around using AI well, openly and on the record, and you shrink the space where cheating even makes sense.
A trap catches a kid once. A redesign changes what the kid does forever.
— Russ
Require every major assignment to state AI's role up front: assumed, permitted with disclosure, or excluded, and why.
Move grade weight onto what machines can't do for a student: oral defenses, in-class writing, and the visible process behind a draft.
Build assignments where students use AI and then critique it, since editing a flawed output is a skill the Madagascar essays prove is missing.
Protect deliberate AI-free work too. Pen-and-paper is a design choice, not nostalgia, when you can say what it builds.
While colleges set traps, a quieter fight is running through K-12, and it starts with a score on a screen. In October, a Palo Alto High School sophomore submitted an essay on The Crucible. Turnitin flagged it as 76 percent likely AI-generated. He was made to rewrite it in class, got a D, and watched his course grade drop. His family assembled 1,162 pages of drafts, notes, and revision history showing the essay being written line by line. The grade stood. In May, his father sued Palo Alto Unified in federal court, alleging discrimination and a lack of due process; the district denies all claims, and the case is unproven and pending. Whatever the outcome, a grading dispute is now a federal docket number.
In North Carolina, it took a 15-year-old to expose the machinery. Green Hope High freshman Eleanor Canina got a zero on a Romeo and Juliet essay she wrote herself. Her teacher ran it through three different detectors and got three different numbers: 62, 75, and 87 percent. On appeal, another teacher reviewed the work and changed the grade from a 0 to a 100. The same essay, by the same girl, scored anywhere from two-thirds machine to perfectly human, depending on which tool you asked. The research says this is not bad luck. Stanford researchers found seven major detectors falsely flagged 61.3 percent of essays written by non-native English speakers, while scoring near-perfect on essays by native-speaking eighth graders. The kids most likely to be wrongly accused are the ones learning English.
The policy world has noticed. Wake County's draft AI policy now discourages detectors outright, citing their "technical unreliability, inaccuracy, and potential for bias", and Oklahoma's new Responsible Technology in Schools Act bars AI as the primary basis for grading or discipline, per FutureEd's tracker.
OUR TWO CENTS
Run the Canina case through your own district's process and see what happens. A teacher gets a score, a kid gets a zero, and unless a parent fights, that is where it ends. The score looked like evidence. It was a guess with a percent sign.
So set the rule before the school year starts: a detector score can open a conversation and can never close one. No grade, no discipline record, no integrity finding rests on a number from a tool whose maker acknowledges it can be wrong. Require human review, multiple forms of evidence, and a look at the student's process, drafts, and history before any consequence. Give families written notice and a working appeal, because the appeal is the only thing that saved Eleanor Canina. And check who your current practice hits hardest; the research says it is your English learners. This is not going soft on cheating. It is refusing to outsource an accusation to a tool that gave one essay three different answers.
— Tracy
Put it in writing: detector scores may prompt review, never consequences on their own.
Require a student conversation and process evidence, drafts and revision history, before any integrity finding.
Build the appeal path and tell families it exists. An appeal nobody knows about is not due process.
Audit past AI accusations by student group. If English learners are overrepresented, your tools are doing the accusing.
The Opposite of Cheating: Teaching for Integrity in the Age of AI by Tricia Bertram Gallant and David A. Rettinger
Bertram Gallant runs academic integrity at UC San Diego, Rettinger teaches psychology at the University of Tulsa, and between them they have spent decades studying why students cheat. Their answer to the AI era is the argument of this issue: stop building better mousetraps and start designing courses where integrity is the path of least resistance. Practical, research-backed, and blunt about the limits of policing. If your leadership team reads one thing before rewriting the honor code, make it this.
Chat with ClassCloud
We’re listening. Let’s Talk!
This newsletter works best when it’s a conversation, not a broadcast. If you want to talk through how any of this applies to your district specifically—or if you have feedback on what would make this more helpful—just hit reply. We read and respond to everything.
Schedule a Virtual Meeting
Thanks for reading,
Russ Davis, Founder & CEO, ClassCloud ([email protected])
Tracy Walters, Director of Customer Growth ([email protected])
ClassCloud is an AI company, so naturally, we use AI to polish our content.
