ChatGPT Detection Tools – neither approved nor recommended at USask
Within weeks of the release of ChatGPT, generative artificial intelligence (GenAI) detectors were claiming to be able to identify the text it produces. However, a growing body of research has found that these detection tools are not effective. Even the company that created ChatGPT (OpenAI) abandoned their detector in July 2023 due to its low rate of accuracy.
Consequently, such tools are neither approved nor recommended for use at the University of Saskatchewan.
Here is what recent research has demonstrated about the lack of efficacy of these tools:
1.) They are unreliable – False positives can lead to false accusations
False positives occur when a detector concludes that text was produced by GenAI when it was not. See Why AI detectors think the US constitution was written by AI to learn why some topics are more likely to be flagged as GenAI output.
Problems with bias
- Those who write English as an additional language are far more likely to be falsely accused of academic misconduct when using detectors. See this study from Stanford University: GPT Detectors are biased against non-native English writers.
2.) They are unreliable – False negatives lead to missed cases
False negatives take place when the detector concludes that text was not produced by GenAI when it was.
- Research on 12 publicly available tools and 2 commercial systems found detectors tend to classify output as human-written rather than detect the GenAI generated text
- Detectors change their settings to avoid false accusations, and then become less sensitive to real cases of GenAI. View this video to see how Turnitin describes why avoiding false positives in their detector can mean more false negatives.
- Vanderbilt University explains here why they turned off Turnitin.com as a GenAI detector.
“Breaking” the detectors is easy.
- YouTube is full of advice of this kind. Try looking up: ways to bypass or trick ChatGPT detectors, to see the many strategies.
- Revising generated responses within ChatGPT to use more sophisticated or varied linguistic patterns can work to avoid detection.
- Applying “a light paraphraser” can “break a whole range of detectors” say University of Maryland researchers.
3.) Their use raises copyright concerns
It is a potential copyright violation to put the work of another person into a third-party tool without their permission. If you, as the instructor, do not acquire permission from each student for submitting their work to GenAI or plagiarism detectors, doing so could be a copyright infringement issue and/or a violation of the University’s Use of Materials Protected by Copyright Policy.
GenAI detectors are not approved nor recommended USask tools.
No detection tool has been approved for use at the University of Saskatchewan.
Staff in Teaching and Learning and ICT continue to monitor developments.
Reach out to the Gwenna Moss Centre team at gmctl@usask.ca for assessment ideas; contact Susan Bens at susan.bens@usask.ca for academic integrity strategies.