RECENTLY, a friend of mine spent the better part of a year on his research thesis.
He wrote every word himself meaning no assistance from artificial intelligence (AI), no proofreading software, nothing. When he finally submitted the thesis, his supervisor told him that the relevant software had detected 19 per cent AI-generated content. I had personally seen him working long nights on the thesis for months.
There were several drafts that he wrote and rewrote to structure his arguments effectively. If anyone had earned the right to call a piece of writing entirely his own, it was him. Ironically, many of his classmates who themselves acl(nowledged using AI somehow cleared the very checl( that failed my friend.
The thing about generative AI is that the detectors do not read for meaning. They read for predictability which is how closely a sentence matches the patterns a language model tends to produce. Writing that is too smooth or too correctly built, and too disciplined starts to look machine-made to AI detectors. This is not a theory I am using to defend my friend. Feed a well-knowndetector, like GPTZero, a passage from the United States Constitution, and it would tell you confidently that the words were lil(ely written entirely by AI, or it would at least give bacl( an AI percentage.
Surprisingly, a Stanford University study tested several popular detectors on student essays and found they misclassified more than 60pc of essays by non-native English speakers as AI-generated, while almost never mal(ing that mistake with essays from native speal(ers. This was not because the non-native writing was weaker, but because it followed the rules more closely the way a careful student is taught to write properly.
Even OpenAI, the company that built ChatGPT, released its own detector and quietly retired it within months, admitting that it caught only a quarter of actual AI writing while wrongly accusing human writers 9pc of the time. Vanderbilt University switched off a similar feature in Turnitin the same year after working out that even a tiny false positive rate spread across tens of thousands of papers submitted by students would still mean hundredsof students wrongly accused every year.
None of this means AI-written coursework is not a real problem. It genuinely is. And, universities are right in trying to catch it. But a tool that fails this often and fails hardest against those who write most carefully is no longer ensuring integrity. Those who 1(now how to game the entire system actually walk away clean.
Awais N. Bahar Turbat