We’ve built this page as a hub for both students and teachers to get some clarity on the abilities and limitations of AI scanners. If you’ve been falsely accused of AI cheating, we invite you to share this page with your educator or bring up some of the background information we provide here.
AI Detection is Not Foolproof
If you’ve seen a legal drama, you’re likely familiar with a scene where a suspect is hooked up to a polygraph or lie detector machine and asked questions. These tests rely on monitoring the physiological responses of a subject, such as heart rate, blood pressure, and respiration, to determine the veracity of the subject’s statements–whether or not they’re lying. These tests have been widely used for decades, but they are not foolproof and can be prone to errors and manipulations. If you give someone a lie detector test, it’s likely that you’ll get the truth out of them, but there’s enough of a chance someone will get away with lying or be falsely accused when they’re telling the truth that these tests are generally inadmissible in court.
While AI detectors are generally more accurate at determining if text was generated by GPT-4 or a similar model, there’s still enough false positives that relying on a scan alone is just not appropriate in the context of serious consequences like accusations of academic misconduct. AI scanners typically analyze text for specific patterns, structures, and inconsistencies that may indicate the use of AI-generated content, but it’s not impossible for a person to naturally produce writing that looks like something made by a language model.
False Positives and Consequences
We scan enormous amounts of text while updating our product, adding scanners to our tool (which, importantly, only uses these scanners as a part of a larger effort to determine a paper’s originality), and we do occasionally see confirmed-human text originating from long before the advent of AI get flagged as AI-generated.
Even a scanner claiming to be 96% accurate can still result in more false positives than true positives, especially when only a small percentage of students are actually cheating with AI-generated text. To put this into perspective, imagine a school with 1,000 students where only 10 students have used AI-generated text. A 96% accurate scanner would correctly identify around 9 out of the 10 cheaters but may falsely accuse 40 innocent students in the process. Clearly, such a high rate of false positives is unacceptable and can have serious consequences for students’ academic careers.
What to do when you’re accused
One suggestion we often see is to have falsely accused students scan the published work of their accuser. If you’re a tenured professor, you’ve almost certainly published enough work for an AI scan to find something that sets it off eventually. Insisiting that a scan alone is enough to determine authenticity would make this immensely embarrasing–and, if the work is recent enough, potentially professionally damaging.
We understand that ensuring academic integrity is a collective effort, and we are here to support both students and teachers in this endeavour. If you are a teacher seeking more information on how to accurately identify AI-generated text or a student concerned about being falsely accused, please feel free to reach out to us. We would be more than happy to provide further assistance, address your concerns, and even communicate directly with teachers if necessary.