Not reliably enough to accuse a student. Detectors give probabilities, not proof. They have flagged honest writing as AI-generated, they perform unevenly across groups of students, and even the companies that make them say the results should start a conversation, not end one.
They can flag work written by real students.
In a Stanford study of seven AI detectors, researchers tested 91 TOEFL essays written by non-native English speakers. More than half (61.22%) were classified as AI-generated, and 97% were flagged by at least one detector. The same detectors were near-perfect on essays by U.S.-born eighth graders.
OpenAI shut down its own detector.
OpenAI's classifier correctly flagged only 26% of AI-written text and wrongly labeled 9% of human-written text as AI-written, on its own test set. As of July 20, 2023, it was no longer available "due to its low rate of accuracy."
Even detector makers urge caution. Turnitin reports a document-level false positive rate below 1% for documents with at least 20% AI writing, and about 4% at the sentence level. It also says there is no "right" AI score and advises teachers to "use the information to initiate a conversation, not to draw a conclusion."
A small rate still adds up. If a school checks thousands of assignments, even a 1% error rate means real students get flagged for work they wrote. A flag can trigger a grade penalty, a disciplinary record, or a damaged relationship with a teacher. And the Stanford findings show errors aren't spread evenly: writers who are still learning English were flagged far more often.
The Center for Democracy and Technology surveyed teachers in 2023-24 and found that 68% regularly used an AI detection tool, and 64% said students at their school had faced negative consequences for using, or being accused of using, generative AI on an assignment. That is a 16-point increase over the previous year. [Source: Center for Democracy and Technology, March 27, 2024; link: https://cdt.org/press/new-research-educators-still-struggling-with-generative-ai-detection-discipline-and-distrust-despite-increased-school-guidance/]
Detectors try to answer "did AI write this?" after the work is done. Assignment Makeover asks a different question before it's assigned: "What would a student have to do to complete this?"
When an assignment asks for personal connection, visible process, specific context, or a live demonstration, there is real student thinking to see, and less to guess about. You don't need a detector to tell you what a student understands if you've designed the assignment to show you.
If you have a concern about a particular submission, talk with the student about their topic, their process, and their ideas.
Should my school stop using AI detectors?
That's your school's decision. If you use them, treat a score as one reason to start a conversation, never as proof, and never as the only basis for a penalty.
Can students get around detectors?
Yes. Rewriting, paraphrasing tools, and editing AI text can lower detection. Students can also prompt a chatbot to "write this so I don't get caught."
Does redesigning assignments make cheating impossible?
No. It makes it harder and less useful, and it puts real student thinking in front of you.