BU Today: The report says that Boston University should “critically embrace” the use of AI. In your view, what does that mean, practically speaking, for faculty and for students, to critically embrace this technology?

Yannis Paschalidis: Acknowledge that AI is here to stay and it can be an effective tool that can be used to accelerate research and enhance teaching. But, at the same time, acknowledge its limitations, outlined in the report, and understand the ethical and sustainability concerns associated with training generative AI models. In practice, this implies using it with caution. With the University adopting a policy to “critically embrace” AI, schools and colleges would be guided to adopt their own local policies adapted to the needs of the disciplines they serve. These policies should be consistent with the University policy but, respecting academic freedom, leave room for faculty to make their own decisions on generative AI policies they adopt for their classes.

Wesley Wildman: The embrace side of this judicious approach to GenAI is rooted partly in the fact that almost every aspect of our students’ careers will be affected by GenAI, and graduates with excellent skills in prompt engineering and flexible, fluid deployment of GenAI tools will enjoy a competitive advantage in the workplace. The critical side involves known and unknown ethical risks. 

On the unknown side, consider that longitudinal studies have recently established that extensive social media use among teenagers leads to deficits in face-to-face socialization skills, including social anxiety and intensified loneliness—an example of unintended side effects of an apparently irresistible suite of technologies. What might the unintended side effects be of extending human cognition to the point that it becomes dependent on GenAI tools? Perhaps no longer knowing how to do simple arithmetic is a tolerable consequence of extensive machine use, but is the loss of generative abilities or reduced creativity—consequences that seem likely—an acceptable consequence of extensive use of GenAI? 

On the known side, we already confront ethically and sometimes legally laden challenges, such as using work of others for training (with an associated IP problem), low-paid workers for feedback-based model training, traumatic exposure of low-paid content moderators, output mimicking styles of human creators, deep-fake audio and video used to spread disinformation (political campaigns, cyberbullying, revenge porn…), violation of privacy and security expectations, economic disruption through job displacement, climate impacts due to the massive server farms that train and run widely used GenAI tools, and inequalities of access. The individual and institutional challenges of the critical-embrace philosophy are to develop necessary skills while keeping eyes wide open to these known and unknown problems.

BU Today: The report also says faculty should use AI-monitoring programs cautiously. It may be fair to say that students are using AI right now. Can you elaborate on why you want to proceed cautiously on this and how you hope to see BU promote academic integrity in the age of AI?

Paschalidis: With respect to programs that detect whether generative AI has been used, we just want to emphasize that they do not provide “bulletproof” answers. They just provide a probability that a certain passage (or image, audio, etc.) was generated by AI. They can also be inconsistent, giving different answers for the same input. Thus, and particularly for assessing academic misconduct, they should be used with caution and understanding their limitations. They are not reliable enough to be used as the only piece of evidence.

Wildman: Ideally, students and instructors can both use the same GenAI-detector apps and get the same probability rating, just as they currently do for plagiarism-checking apps. At present, however, GenAI-detector apps appear to be inconsistent, in the sense of sometimes giving different ratings at different times on the same piece of prose. They are also unreliable in the sense of being inaccurate, with different detectors being easier to deceive when students introduce spelling or grammar errors or modify GenAI prose specifically to avoid detection. Accordingly, we are advising instructors and students to stay apprised of detector technology developments and to exercise caution and explicitness in their use. It is also important to notice that some companies appear to be cashing in on the established reputations of existing detectors (e.g., the new arrival ZeroGPT is being confused with the original GPTZero, though the two are very different in reliability).

BU Today: Can you talk about the timing of this report: why BU undertook this effort, how long it took, why it’s being released now, and when you hope to see some of the recommendations implemented?

Paschalidis: The taskforce was formed in the fall semester and met throughout fall 2023 and early spring 2024. We heard from many experts, consultants, people in industry, and many experts we have at BU. The roster is representative of the many disciplines we have here, and each discipline brought a different perspective on generative AI. The motivation for the task force is, of course, that AI is revolutionizing research, teaching, our daily lives, and rather than having a patchwork of local policies it is prudent for the University to have a unified policy framework and devote resources to enable training and support and incentivize new courses and new ways of teaching existing courses.

Wildman: We heard from experts representing disciplines not already named to the task force. We hope to see key recommendations implemented immediately, especially the “critical embrace” posture. Some elements will take longer, as they involve infrastructure and personnel investments.

BU Today: Can you talk about the risks versus the rewards of AI technology in higher education and how BU hopes to strike that balance?

Paschalidis: In terms of research, some of the risks are associated with uninformed use of the technology (e.g., not understanding the risk for hallucinations) and its potential to “turbocharge” and make it harder to detect plagiarism. In terms of teaching, some of the risks include overreliance and atrophy of critical skills. Hence the need to critically embrace [see first question above]. Beyond recommending AI literacy for everyone, the report includes recommendations to help mitigate these risks: for instance, adapting class assignments to emphasize process over the final product and guide students through an iterative process to produce a final product, whose steps could involve using generative AI.

Wildman: Turbocharging plagiarism refers to asking GenAI to paraphrase an existing block of text written by someone else, and then including the paraphrase as one’s own writing without citing the original source or acknowledging the role of GenAI in producing the paraphrase.