Generative AI and Academic Honesty
Detecting AI - Embracing AI.
Students using Generative AI tools, such as ChatGPT, to cheat is a major concern among educators and educational administrators. For example, a student could use ChatGPT to generate an entire essay and hand it in as theirs. Can AI writing be detected, and is AI writing always undesirable? Let’s dive into it.
Several online AI detectors have been made available recently to assist teachers and professors in identifying cheating with Generative AI, indivatively GPT Zero and Turnitin’s AI Writing Detection tool. In a plot twist, such tools might just don’t work, as one can read in numerous sources, including articles in MIT Technology Review, Washington Post, and Virtualization Review. To make things worse, there is evidence that AI detectors might be biased against non-native English writers (writing in English), according to Stanford University researhers. Therefore, educators and administrators might need to consider disregarding AI detectors altogether.
Let’s investigate a simple, relevant scenario.
A teacher assigns essay work to be submitted on the course’s learning management system, requesting “originality” (no Generative AI allowed). A few students cheat by delivering ChatGPT’s content. The teacher passes all submissions through an AI detector.
We consider the binary case of each student’s work to be either AI-generated or non-AI-generated, and the detector’s output to be one of those two outcomes. Trying to solve binary classification problems like this, researchers are using “confusion matrices” that summarize their model’s performance. For our conceptual, qualitative scenario, a possible confusion matrix could be the following.
Note that “positive” and “negative” correspond to the AI detector predicting AI and human work, respectively, while “true” and “false” correspond to the detector’s performance in relation to the actual work (successful and unsuccessful, respectively).
The “fun” part: The educator can only hypothesize about the true nature of each student’s work! Apparently, there are four different possible actual work vs. AI detection states, colorcoded in the confusion matrix as either green (desirable) or red (undesirable):
- True positive: The detector has identified academic misconduct.
- False positive: The detector’s prediction of academic misconduct is inaccurate, therefore the student might be falsely accused for cheating.
- True negative: The student has taken responsible action per instructions. This is the ideal scenario — and probably the most common situation teachers and professors might find themselves in.
- False negative: The detector fails to recognize AI-generated work, therefore academic misconduct might remain unnoticed (given AI detectors are the only tools employed to identify cheating).
Consequently, even if AI detectors were reliable, the educator could not just count solely on their predictions anyway. Moreover, the binary case above does not accurately represent the emerging reality of students using Generative AI at some extent (neither 0% nor 100%) before they finalize and submit their work. Furthermore, there are many fruitful ways that Generative AI could be used by students, such as to brainstorm ideas or to edit a draft manuscript.
But will chasing down the use of AI benefit the students’ learning?
Generative AI is revolutionizing all sectors of the economy. The vast majority of today’s students will be “collaborating” with Artificial Intelligence for their entire social and professional life. In the near future, the majority of the students will be AI-natives. Naturally, academic institutions should be preparing students to live and work with AI. Among many other interventions, students should be encouraged to use Generative AI to augment their learning — to collaborate with AI and their classmates for problem solving in a “Mixed Intelligence” context.
Still, educators will need to plan for AI-related academic misconduct.
Practically, the educators could 1) clearly communicate assignment instructions, relevant AI policies, and overall expectations to their students regarding work that could be potentionally conducted by Generative AI, and 2) design new or modify existing assignments in accordance with those expectations (in reverse order). Do we want human-only work? Let’s assign essays during class, assess orally, require higher-order thinking, etc. For mixed intelligence work, let’s set boundaries regarding the AI part of the partnership, require more complex work in shorter time, request the prompts used by the students, etc.
Securing both instructional design and communication, the risk of confusion and/or misconduct by the students will remain low, as the decision map below suggests.
In terms of addressing academic dishonesty, and apart from considering the design and communication of assignments, academic institutions could indicatively strategize towards addressing the root cause of academic misconduct, promote Ethos, educate students about AI, train professionals in AI, and employ learning analytics (more in my article here).
How much learning space will we secure for human-AI collaboration?
