On this page
Some schools and universities run student essays through tools that claim to spot AI-written text. If you have ever been told your own work "looks like AI", you are not alone, and the research suggests the tool may be the problem. This post explains how these detectors fail, who they fail most often, and what you can do to protect honest work.
How accurate are AI-writing detectors?
Not very. The clearest sign came from OpenAI, the company behind ChatGPT. In January 2023 it released a classifier to flag AI-written text, and reported that on its test set it correctly identified 26% of AI-written text while wrongly labelling human writing as AI 9% of the time. On 20 July 2023 it withdrew the tool, saying it was no longer available "due to its low rate of accuracy".
Independent testing found similar problems. Weber-Wulff and colleagues (2023) tested 12 free tools and two commercial systems, including Turnitin, on documents written by people, by ChatGPT, and by ChatGPT then edited or paraphrased. They concluded the tools were "neither accurate nor reliable". All scored below 80% accuracy, and only five scored above 70%. Simple editing or paraphrasing of AI text made detection much worse.
Why do detectors flag human writing?
Many detectors look at how predictable the words are. AI text tends to use common words in common patterns. The trouble is that plenty of human writing does too, especially careful, simple writing by someone working in a second language.
A Stanford team led by Liang (2023) tested seven widely used detectors on 91 essays written for the TOEFL English test by non-native speakers, and 88 essays by US eighth-grade students. The detectors judged the US essays accurately. But they labelled the TOEFL essays as AI-generated with an average false-positive rate of 61.3%. About one in five TOEFL essays was flagged by all seven detectors, and 97.8% were flagged by at least one. When the researchers used ChatGPT to make the TOEFL essays' vocabulary fancier, the false-positive rate fell to 11.6%.
For many Indian students who write in English as a second or third language, that finding matters. Plain, correct English is not evidence of AI use.
Why do small error rates add up?
Even a detector that sounds accurate will wrongly flag many people when it checks a lot of work. When Vanderbilt University turned off Turnitin's AI detector in 2023, it did the arithmetic. Turnitin had claimed a 1% false-positive rate at launch. Vanderbilt had submitted 75,000 papers in 2022, so about 750 papers could have been wrongly labelled as partly AI-written.
Try it for a class. Suppose 60 students all write their own essays and a detector with OpenAI's reported 9% false-positive rate checks them. You would expect 60 × 0.09 = 5.4, so about five honest students flagged. At 1% the number drops to below one per class, but across a whole school year of assignments it climbs again.
What should students do to protect their work?
Keep your process visible. Write in a tool that saves version history, such as Google Docs or Word with autosave, and do not paste the final text in from elsewhere. Version history shows the essay growing over time.
Save your rough work. Outlines, handwritten notes, early drafts and the sources you read are strong evidence of your own thinking.
Know the rules. Read your school's or university's policy on AI. Some allow AI for brainstorming or grammar checks and ban it for drafting. Our post on AI and academic honesty covers where the lines usually fall.
If you are flagged, stay calm and ask for the evidence. Ask what tool was used and what score it gave. Offer your version history and notes, and offer to explain your argument or redo part of it in person.
Do not try to game the detector. Running your work through a paraphrasing tool to avoid a flag can itself break the rules, and it weakens your case if questioned.
Using AI without losing your own voice
The safest way to use AI for schoolwork is in a way that leaves your own reasoning on the page. Fermi's AI tutor teaches rather than answers: it decides whether to explain, ask or show, so the work you hand in is still yours.
Before your next assignment, turn on version history in your writing tool and keep your outline in the same file.
