Fraud / Generative AI (Chatgpt)

Case description

In a written test, students A and B submitted solutions contained parts apparently not written by themselves, judging by the hand writing. Some answers on A’s sheet were written in B’s hand, and vice versa for other questions of the same exam. Moreover, the answers, though not identical, showed strong correspondences, including atypical errors made in both cases. Apparently the students had prepared this fraud together, each of them made part of the exercises twice during the test (once for him/herself, once for his/her partner), then swapped answer sheets when handing in their results.

Proceedings

When confronted, the students denied their actions. The written tests were then submitted to an external handwriting expert, who concurred with the finding of correspondence between the two sets of handwriting. When presented with this independent opinion, the students admitted their fault, confirming the procedure they had followed (described above).

Outcome

This was regarded as a severe fraud case, because the students had planned it carefully and agreed to act fraudulently (rather than acting on the spur of the moment); and also because they at first denied their actions when confronted with the evidence. 
Both students were barred from taking any examination in the next quarter, and warned that they would be evicted from the programme in case of a repeat infraction.

Case description

Student A, upon failing to meet a deadline for an individual take-home test, copied the solution of student B, who had been careless enough to put it on a publicly accessible repository, without knowledge of B. This was detected; A admitted his fault immediately, upon which the case was referred to the Examination Board.
Before the Examination Board had come to a verdict, the same student A, now working together with another student C on another project for the same study unit, repeated his plagiarism by again copying solutions from an open repository of student D. This was done without knowledge of either C or D. When confronted with the evidence, A once more immediately admitted his actions, claiming full responsibility.

Proceedings

Because A admitted his fraud immediately, the facts of the matter are clear. The course unit reader contains a description of what constitutes fraud that includes exactly the case of copying from an open repository. 
Students are warned not to put solutions to tests and projects on an open repository, with the explicit statement that doing so may make them accessories to this type of plagiarism. Awareness of this is, however, still low.

Outcome

Although A admitted his fault the first time around, he clearly still did not understand or appreciate the severity of his actions. For the first instance of fraud, the sanction was to declare the test results invalid, meaning that he could not pass the study unit any more. For the second instance of fraud, the sanction was to bar the student from taking any examination during the next quarter.

‘De eerste AI-paper stond voor de rechter, en verloor' - ScienceGuide  
In the Science Guide an article about the first AI case before (de Afdeling van) the Raad van State (but it was some time ago, so there may have been more since then). The article is only available in Dutch. It is interesting because of the criteria used and Van Berkel's challenges. Which leads to an interesting question: Can we detect and prove the use of AI without errors in the work (e.g., incorrect references)? 

The gist of the story: The Raad van State concluded that it has been established beyond a reasonable doubt that a student improperly used AI, and the appeal of the student (against a previous ruling by the CBE) was declared unfounded. It rejected the student’s claim that the issues were merely careless mistakes, considering the number and nature of the inaccuracies and the relatively small scope of the assignment. A key aspect of the decision was the criteria used to determine fraud. The evidence was indirect: no one directly observed the student using AI. Instead, the judgment relied on circumstantial evidence — facts that do not directly prove misconduct but allow it to be inferred. 

According to established case law, two conditions must be met when relying on indirect evidence:  
1) A single indirect fact is usually insufficient.
2) A conviction must be based on a combination of circumstances that together leave no reasonable alternative conclusion than that the person committed the violation.
Together, these requirements are referred to as a “closed chain of evidence.”

Van Berkel (previous UHD Universiteit Maastricht and assessment expert) challenges the ruling. While the first criterion has been met, it is not clear that the second has been satisfied. The judgment mainly relies on the errors attributed to AI (e.g., incorrect references). However, why are these considered typical AI errors? If a paper is flawless, does that mean no AI was used? According to Van Berkel, the conclusion is somewhat problematic.