Detection of GAI tools

Thinking about rules and red lines around examination would be easier if we could identify work generated by GAI tools, but this is not possible, and may never be. Despite the rush to market products to detect the use of GAI tools, nothing has yet been shown to be reliable, and the University does not recommend their use for any purposes (except perhaps researching their effectiveness).

The products released to date have been shown to be likely to lead to both false negatives, where the text was created with a GAI tool, but this isn’t detected, and false positives, where the detector says the work was created with a GAI tool when it wasn’t  (Weber-Wulff et al., 2023 Links to an external site.). The second of these is particularly concerning, because it means that students could be accused of using a GAI tool to produce their assignments with no reliable way to prove guilt or innocence. Worse, the detectors are more likely to misidentify writing in English from non-native English speakers are AI generated, even if it isn’t (Liang et al, 2023 Links to an external site.). Perkins et al Links to an external site.(2024) found that deliberately manipulated texts made the detectors even less accurate. ChatGPT quietly withdrew its own detector tool Links to an external site. in July 2023 because it is unreliable. If you would like to read a straightforward analysis of why it is likely that we will never have a good tool to detect GAI use, I recommend this article Links to an external site. by Benj Edwards (2023).

Lund University does not subscribe to any GAI tool detection software and does not recommend that detectors are used. You should not upload student work to any such tools unless they have been verified for data protection and accuracy. In module 3, we will look at other ways to manage student assignments which assure assessment validity and reliability when GAI tools are available.

Just in case you were wondering, we teachers are also no good at detecting AI use, even if we might think we have these special skills (Fleckenstein et al, 2024 Links to an external site.), although there were mixed findings with a small study of peer reviewers vs. AI detectors reading scientific papers (Liu et al, 2024 Links to an external site.)

Summary: there is no reliable way to identify work which has been generated by a GAI tool.