AI cheating tools outsmart detection software as schools face impossible enforcement battle

AI cheating tools are winning. Detection isn't the point.

A wave of apps now rewrites AI essays and types them out with believable typos, beating the software meant to catch them. But the detection arms race misses the real story: education long ago made the grade the goal, and AI is simply the most efficient optimiser yet. The videos are everywhere, and the offer is always the same. Let AI do your homework, and you will not get caught. According to a New York Times investigation, TikTok and YouTube are now full of tutorials selling students two kinds of tool. Humanisers rewrite AI-generated text so it no longer reads like a chatbot. Autotypers do something sneakier: they drip the words into a document over hours, faking typos, deletions and edits so a finished essay looks like a real writing session. Both are built to defeat the software teachers use to catch AI. The same companies sell the disease and the cure Here is the uncomfortable part. Some of the firms selling detection tools also sell the tools that beat them. Grammarly, now owned by Superhuman, offers teachers an "authorship" checker that scans a document's history for signs of AI. The same app will also generate text from scratch, "humanise" it, and rewrite phrases that might trip a detector. GPTZero, a detector born as a Princeton thesis, can also write a full paper, complete with citations, in seconds. The NYT found a marketer had built a fake teaching-assistant persona on TikTok to push it to students. Jenny Maxwell, who runs education at Superhuman, was blunt about where this leads. The race between detection and evasion is, she said, "ultimately, a dead end." Her summary: "Bigger cat, bigger mouse." The detectors barely work anyway She has a point, because the cats are not very good. University of Florida researchers tested the five most popular AI text detectors and found false-negative rates as high as 99.6 per cent, with a single vocabulary tweak enough to defeat most of them, Digital Trends reported. The tools also throw false positives, disproportionately flagging non-native English speakers. So schools that discipline students on a detector's say-so are standing on very thin ice. The technology they are trusting is, by its makers' own admission, losing. From oral exams to internet blackouts Faced with that, institutions are improvising, and the responses range from sensible to extreme. At the calm end, Harvard professors are leaning harder on oral and pen-and-paper exams, which a chatbot cannot sit for you. At the other end sits coercion. To stop cheating in its national medical-school entrance exam, India ordered Telegram blocked for several days, The Register reported, after the test was annulled and rescheduled following a suspected leak. More than two million people sit that exam for roughly 100,000 places. Digital rights groups called the shutdown disproportionate, and it is part of a wider pattern of governments cracking down on AI misuse with very blunt instruments. The number was always the problem Step back, and the cheating panic looks like a symptom of something older. School turned learning into a single number, the grade, a long time ago. The philosopher C. Thi Nguyen calls this "value capture": you adopt an external metric, then let it quietly replace the thing it was meant to measure. In his book "The Score", reviewed this week by MIT Technology Review, he points to the GPA as the classic case. Students stop chasing understanding and start chasing the grade. It is Goodhart's Law in a backpack: when a measure becomes a target, it stops being a good measure. AI is just the most efficient optimiser yet invented for that target. If the point of the essay is the score, not the thinking, then offloading the thinking is the rational move, even as studies warn that this kind of cognitive offloading lets real skills wither. A gas pedal, no brake The people building this technology are uneasy too. Anthropic co-founder Jack Clark told the BBC the industry "has a gas pedal, but it doesn't have a brake pedal", and noted that Anthropic's own model now writes most of its code. His company has called for a coordinated brake on frontier AI. Maxwell, from the other side, argues that withholding AI from students is "educational malpractice", since they will use it at work regardless. Both things can be true. The detection arms race cannot be won, and detection was never the real question. The harder one, the one schools have dodged for a century, is what the grade is actually for. AI did not create that problem. It just made it impossible to keep ignoring. Until someone answers it, the bigger cat will keep chasing the bigger mouse, and the mouse will keep winning.

AI tools that help students cheat are multiplying, and the detectors can't keep up

A New York Times report has found that cheating tools are evolving faster than the software meant to catch AI writing. A wave of new apps marketed on TikTok and YouTube is making it nearly impossible for teachers to tell whether students are actually writing their own homework or offloading it to AI. The New York Times reports that tools known as humanizers and autotypers have closed the gap that used to give AI-written homework away, and that the same companies selling detection software are sometimes the ones helping students get around it. The tools work around the checks teachers rely on Humanizers take AI-generated text and rework it so it no longer sounds robotic or repetitive enough to trigger detection, while autotypers solve a timing problem. Instead of a thousand words appearing in a document all at once, which can tip off a teacher checking version history, autotypers release the text gradually over hours and even insert fake typos, deletions, and edits to mimic a real writing session. Apps like Dripwriter and Duey.ai advertise this directly, telling students they can step away entirely and still turn in something that looks self-written. One app, called Typeflo, promised students could relax and eat a sandwich while it produced their essay. It turned out to be built and marketed by the teenage son of an Emory University professor, who said he hadn't known the extent of its social media presence and pulled it down after being contacted. Even the detectors built to catch AI can't be trusted GPTZero's entire pitch rests on detecting AI writing that other tools miss, but the Times found that a marketer paid by the company had built a fake graduate teaching assistant persona on TikTok to promote it to students. The videos walked students through GPTZero's browser extension, showing them how to screen a paper for AI flags before submitting it and revealing that the same tool could generate a full paper with citations from scratch. Responding to the report, GPTZero's co-founder and chief executive, Edward Tian, said the company has cut ties with the marketer and is reconsidering whether to keep that paper-generating capability. Grammarly faces a similar contradiction, offering an authorship checker for teachers while also providing a humanizer, text generation, and paraphrasing tools on the same platform. That unreliability isn't limited to these two companies either. Recommended Videos A report from earlier this year revealed how University of Florida researchers tested the five most popular AI text detectors and found false negative rates as high as 99.6 percent, with a single vocabulary tweak defeating most of them entirely. The findings suggest that schools leaning on these tools for disciplinary decisions are working with far less certainty than they assume. Outlawing AI in classrooms might sound like the obvious fix, but with detection this unreliable, schools may have no way to enforce it even if they tried. Some educators argue that's beside the point anyway, since students will need these same tools the moment they enter the workforce.