Scholarship in the Age of AI
While academic institutions are scrambling to pretend that they are at least partially in control of the AI revolution, some more established actors have already had to adopt new measures. The venerable arXiv, for example, recently announced that they will ban authors for a year under certain circumstances:
If a submission contains incontrovertible evidence that the authors did not check the results of LLM generation, this means we can’t trust anything in the paper. The penalty is a 1-year ban from arXiv followed by the requirement that subsequent arXiv submissions must first be accepted at a reputable peer-reviewed venue. Examples of incontrovertible evidence: hallucinated references, meta-comments from the LLM (“here is a 200 word summary; would you like me to make any changes?”; “the data in this table is illustrative, fill it in with the real numbers from your experiments”)
I believe that the policy is right in spirit but I am worried about the way it will be imposed and whether there is due process for all. That being said, in this post, I am more interested in a high-level discussion on what it means to follow the norms of scholarship in the age of AI.
In a nutshell, my position is this: You are ultimately responsible for your work and cannot abdicate that responsibility to a tool.
The moment we start assigning personhood to an AI system, it becomes a proper coauthor and collaborator, requiring direct credit authorship. But until we are there,1 you cannot shirk your duties. It is perfectly acceptable to use whatever tools you have at your disposal to write your papers, but you are supposed to remain in charge. If you use AI for literature search, for instance, you need to check the results for (a) existence, (b) correctness, and (c) content. You need to do this because scholarly work relies on citations the way good detective work relies on a chain of evidence. You, as an author, have the duty to tell your readers about the relevant literature landscape. It needs to be clear to what extent you are extending the state of the art, relying on earlier models or arguments, and so on. The existence of AI does not change this fundamental prerequisite of scholarship.
Mistakes can happen and will happen,2 but not all mistakes are equal. For example, referring to a non-existent work is highly problematic. It erodes trust in your own work and, beyond that, the trust of the public in science itself—at least to some degree. Referring to claims in a work that are not part of that work is similarly problematic. You are misleading readers while also misrepresenting the work of others. By contrast, referring to a preprint instead of the published version of a paper is, ultimately, harmless. The discourse around the aforementioned arXiv policy misses this type of distinction, unfortunately.
Enough about bibliographies, though! Scholarship is much more than that, but the same principle applies: You are ultimately responsible for your work and cannot abdicate that responsibility to a tool.
Here are some more concrete examples:
-
Using AI to (re)write your paper implies that you need to understand editorial suggestions before accepting them.
-
Using AI to (re)write your code implies that you need to defend or justify modeling choices.
-
Using AI to (re)write your proofs implies that you check their correctness.
None of these examples ask whether it is a good idea to use AI for these purposes. Since I lack concrete data, we are now entering deeply speculative and personal territory. I am going to start with a confession: I derive most of my enjoyment from mulling over things and solving problems. Whether it is writing, reading, or coding, I just love the process as such. Offloading certain tasks to AI robs me of that joy; Terence Tao used the following analogy in a recent Atlantic interview:
AI tools are like taking a helicopter to drop you off at the site. You miss all the benefits of the journey itself. You just get right to the destination, which actually was only just a part of the value of solving these problems.
Since I do not know what, in the words of Marie Kondo, “sparks joy” for you, I can only leave you with the generic piece of advice that you need to decide when to take the helicopter and when to hike yourself. However, when you do take the helicopter, make sure to (a) acknowledge it and (b) check that it actually put you where you wanted and needed to go in the first place.
We are all figuring things out in these times. Have courage to be truthful to yourself, and the rest will follow.