The lack of reliable detection metrics, coupled with genuine uncertainty or mistaken belief on the part of students as to whether they have complied with university or classroom policies, has sent many instructors back to paper-only exams both in general and, anecdotally, at our own university. We believe this teaching approach is profoundly mistaken.

We are not the first to propose proactive engagement with ChatGPT and similar tools, but what we propose is slightly different. In general terms, our suggested protocol contains the following elements:

1. 1+ open-ended question asking students to cite to primary and secondary sources
2. Instruction to paste the original prompt (#1) into a free gen AI tool
3. Questions about the AI response for the student to answer on their own

To build questions to prompt students’ own narrative creation in #3, we suggest four categories of questions:

1. Category I: Find clear deficiencies (if any) in the AI response (factual errors, subject matter not relevant to the question asked, hallucinations, mis-citations, misattributions, etc.).
2. Category II: Critically analyze the AI text. 
   1. How superficial or incomplete is the AI-generated response? 
   2. When using tools with data ingest cutoff dates, are there any recent facts or circumstances that are unaccounted for, including current events or recent scholarship?
   3. What arguments or other readings did it leave out? 
3. Category III: Supplement the response with your own work.
   1. Read additional online references in a physical library or e-library resource (including from primary or scholarly sources), and include links to their listings on our University’s library website.
   2. Provide an additional self-written response on how additional source material enhances student understanding of the original question asked.
   3. Encourage or require students to carry out their own follow-up, perhaps through iterative prompts to the original question and AI-response (with both prompts and AI responses provided to the instructor).
4. Category IV: Personally reflect on the exercise, focusing on how best to use AI-generated sources of responses in future assignments 

These four categories represent our rough attempt to mitigate, if not exploit, known flaws in AI-generated responses to the benefit of advancing critical thinking in learning assessments. The categories can be simplified or extended, and we recognize that the protocol has limitations, especially because it places an extra burden on instructors while they review essays. But the most important goal in this exercise is to encourage students to develop strong AI literacy and critical thinking skills. Especially in this age of mis- and disinformation, such skills are essential for information students to develop. 

Despite our sharing this philosophy with both classes, some students still tried to use gen AI in unpermitted ways. To detect inappropriate gen AI use in both years, the class TA (with assistance from a grading assistant) preemptively generated responses to the essay assignments and compared student submission to those materials. When students had used gen AI inappropriately, common patterns included: introductory and concluding paragraphs that were similar to the AI-generated ones, hallucinated or otherwise inaccurate or problematic references coupled with the lack of short form in-text citations (both of which we required), and repeated misinterpretations or misattributions of quotations or cited ideas. In 2025, we placed a particular emphasis on accurate citations and quotations. This was a time-intensive process, but it is one we believe is worthwhile; without it, students are unaware of the places they make mistakes while using gen AIs in writing. 

Using this protocol, between 2024 and 2025, we achieved a reduction from 99 infractions (2024) of our gen AI policy to 52 infractions (2025). While this reduction was not as dramatic as we’d hoped, it was significant (52.5%). To us, the results signify that, while gen AI is here to stay, our teaching methods reflect that mitigating its improper use is achievable. 

Our primary takeaway is that students need better guidance around source and citation literacy so that they are able to check the sources and citations gen AIs provide to ensure they are accurate. Without those skills, students will likely continue to include erroneous citations and quotations, which will result in continued infractions—both in our class and against university-level academic integrity policies. As such, instructors need to emphasize source and citation literacy in their teaching, especially when using gen AI in their instruction.    

We have proposed one possible solution to gen AI’s challenges to critical thinking. Critiquing gen AIs’ outputs, regardless of version, is an approach that, while time consuming, will continue to be successful going forward. First and foremost, we believe that using this approach has the potential to richly enhance students’ critical thinking skills by asking them to more deeply engage with the material being covered. Secondarily, following the suggested approach broadens students’ understanding of gen AI tools, which helps them develop AI literacy skills. In a world where university graduates will increasingly be expected to be familiar with and use such software over the course of their careers, the third effect of adopting this approach will be to introduce prompt engineering skills, which will likely be desirable going forward.