The arrival of new technology inevitably comes with systemic pressures to adopt, rather than prohibit, its use. When it comes to the use of large language models (LLMs), which I’ll call the ‘common chatbots,’ adoption comes with the familiar challenges: there is no sure-fire way to detect chatbot-generated text, and many detectors have unacceptably high false-positive rates, so there is no principled way to evaluate our students on the merits of their work. To work around this, I’ve recently seen some teachers rule that students can use the common chatbots in their written work, so long as they provide in-text citations to the contributions which originate from the common chatbots. But how does one cite a chatbot? Better yet, can a student cite the common chatbots without undermining the generally accepted purposes of citation practices?
A citation is supposed to give credit where credit is due and ensure that the author of a paper is not passing someone else’s ideas off as their own. Granted, if a student is citing a common chatbot in their paper, they’re not trying to pass the bot’s work off as their own. The question to focus on here, then, is where exactly is the credit due when a student cites a common chatbot? Though some of the more advanced “thinking models” can point to specific portions of its training data which are themselves citable pieces of work, these models are expensive. Regardless of other ethical issues, it would be unjust to form a teaching policy around a thinking model that only wealthier students can access. So, with the common chatbots, the problem is this: as multiple ongoing lawsuits plausibly claim, the basic architecture of the common chatbot is essentially a sort of plagiaristic black-box. It’s always imitating some string of text that the transformer architecture was trained on, but we do not know which strings of text (or the identify of their original authors). Upon close inspection, it seems that the two “AIs”—artificial intelligence and academic integrity—are fundamentally irreconcilable.
Let’s say I construct a special blender which can shred and reform paperback books into novellas. Paper goes in, press blend, and a cut-and-paste, serial killer style letter gets spit out on the other end. If every single book I toss into this hypothetical woodchipper is in the public domain, then this is probably fine. But it would be foolish to think that AI models are only being trained on the public domain. They are being trained on every little bit of data that every big tech company can get their hands on. Yours, mine, and all our students too. Since the American land of the lawsuit is also the land where settlement sums are the ‘cost of doing business,’ intellectual property lawsuits will not necessarily be able to solve this issue. Many big tech firms are already settling to avoid going to trial. One could speculate that such lawsuits stem from a desire to avoid the discovery process, which could reveal potential violations of European laws such as the GDPR.
Some might claim that LLMs are simply the new Wikipedia. But this is misguided. It’s fairly intuitive to integrate Wikipedia into standard research practices. Follow the sources that the Wikipedia article cites and use those instead. No expensive neural network required. The common chatbots are quite similar to Wikipedia in a general sense: they might be an okay starting point for an undergraduate research project, but university-level academic research requires going beyond a tertiary source in one’s readings. One could make the case that locked and heavily moderated Wikipedia articles are actually more reliable than some of the common chatbots, though advances in transformer architecture do continue to lower the rates at which LLMs hallucinate falsehoods. The point is, common chatbots are still late-90s era Wikipedia: we don’t have sufficient norms set up for citing them in written academic work. Since tech companies are still largely reliant on expert input for refining subject matter models, it’s not obvious what ‘peer review’ would even look like for LLM usage in language-based disciplines like the qualitative social sciences and humanities.
To make matters worse, the firms developing such models have no vested interest in peer-reviewed levels of academic integrity. A prohibition on the usage of such technology need not reflect a distrust of the students who might use them. Rather, I do not trust the technology itself. Perhaps that could change in the future. But for the now, instructing a student to ‘cite’ a common chatbot would be only a few steps removed from allowing them to ‘cite’ a purchased paper from another student. It’s not that any usage of artificial intelligence automatically demonstrates a lack of academic integrity. But if we, as teachers, still care about the latter kind of ‘AI’ then let’s hold off on allowing students to use the former kind until the tumbleweeds settle in this digital wild west.
Obviously, many will object on the grounds that we cannot truly enforce such prohibitions for reasons related to the difficulties mentioned above. Moreover, one might even object that such prohibitions could set students up for failure as common chatbots become more integrated into various workplaces. But so far, the evidence suggests that extended usage of the common chatbots for writing and verbal communication has negative ‘cognitive offloading’ effects on the capacities of students. More importantly, younger generations seem strangely aware of these effects and wary of the technology despite being disproportionately likely to use it. This suggests our students are experiencing at least some minor AI-akrasia, if you will. Thus, even if we cannot police common chatbot-usage or fully mitigate its impact on academic integrity, we can rightly point out that students have strong practical reasons to abstain from the common chatbots, reasons which they themselves often recognize. On balance, then, both students and teachers of philosophy can simultaneously prioritize academic integrity and the cultivation of cognitive capacities by setting norms against common chatbot usage.
Isaac Shur
Isaac Shur is a philosophy PhD candidate at Northwestern University interested in ethics, politics, philosophy of technology, and social ontology. Their doctoral research applies novel arguments from value theory to issues of technology, especially artificial intelligence. When not philosophizing, Isaac enjoys cycling, karaoke, and playing games with their friends.
