待翻譯：What Guidance Do We Give Junior Colleagues and Peers on Use of AI?

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

In my sphere there have started being a number of new conversations about use of GenAI in the use of scholarship and how that will impact assessment for promotion and tenure. We’re having these conversations from a “what advice do we give as a college to each other and to our junior colleagues.” It has also come up recently in the emails exchanged for the two external reviews I have been requested to provide this summer. I have explicitly asked for what guidance was provided to the candidates and asserted at least for myself that I won’t put the candidate materials into Co-Pilot, which is something an external reviewer at my institution did last year. The number of colleagues across disciplines who are speed running and openly delighted at describing how they are offloading tasks that are the core of what we as faculty are privileged to do — creating new knowledge, writing up our findings, sharing our knowledge. There’s always been the “oh and I have to teach” aspects of it for many programs that are training researchers but now there’s also this seeming desire to not do their own research too? It’s unclear what they anticipate their job is to be. Anyway, an internal colleague asked for some thoughts related to what I was thinking about as we face cases where junior faculty are navigating the weird pressures of “you must AI OR DIE” versus “so you’re offloading your critical thinking to an ethically corrupt and ecologically disastrous auto-complete calculator.” Here’s a slightly edited version of what I sent her — with a recognition that this is not a comprehensive document but questions I am noodling on… We need to define “AI” and stop lumping everything under a single umbrella term. Are we talking about use of spell check and grammar check that catches the wrong their/they’re/there; testing novel machine learning or algorithmic tools that are working across datasets too large to identify correlations; automating some data clean up and standardization; or generating entire literature reviews and inappropriate rat visualizations? There are significant differences in perceived usage varying from “My campus version of Office 365 has spell check and won’t let me turn off autocomplete suggestions” versus “Here is this manuscript, can you tell that I used an extruded text from Co-Pilot for the literature review and I had it “do” the analysis?” The Artificial Intelligence Disclosure (AID) Framework by Karl Weaver may be useful here. Faculty will need to put in more effort with their manuscripts if they plan to use an LLM. There is no reason that mentors, peers, peer reviewers, editors, or other colleagues should be asked to put in time and effort reviewing and providing feedback on extruded text manuscripts that you couldn’t be bothered to write, particularly when those manuscripts are half-baked. I’m seeing this with students and it’s creeping into peer review. You are not too busy or important to meaningfully engage in drafting your own manuscript. The often very obvious use of LLMs to draft literature reviews suggests a lack of curiosity and engagement with the discipline and your colleagues – if you aren’t reading and engaging with your peers, why are you producing materials that you expect them to read? How can you be sure this research actually meets a gap in the literature? How are you critically engaging and creating a story to lead to your question rather than writing slightly shorter sentences than what is in the abstract? It suggests a misunderstanding of the point of a literature review that is not flattering. Use of LLM tools is likely to reinforce citation biases and has a strong potential to enhance erasure of women, faculty of color, and other minoritized groups in the literature across disciplines. The tools are trained on what is available to them and we know there are significant problems there. The answer cannot be to steal more of our currently copyrighted materials to train a tool that is then sold back to us. The LLM companies are also ingesting retracted papers without capturing the metadata of retraction, which will cause problems where someone aggressively asserts something that has been refuted. Systematic reviews and meta-analyses will get worse and less trust-worthy. This is on top of the disappointing behavior of researchers not bothering to verify that a citation even exists before submitting it to a journal. The role of tenure and scholarship at our institutions is to create new knowledge. I was given tenure not only on what I had accomplished to date –which absolutely mattered — but because the university believed I would continue generating new knowledge and work that impacted my discipline, the community, and the world. We’re already hearing about how the voice of scholarship is shifting to sound like the popular LLMs, the ideas are becoming homogenized because of the reliance on these tools. If everyone is asking the generally same prompt — as evidenced when the students all had an LLM create a personalized near-identical apology for their professor — we are not gaining new knowledge and ideas or developing new voices. LLM tools cannot be authors – as is in the COPE Position, and use of LLMs must be weighed against ICMJE or similar mechanisms to determine whether a faculty member meets the qualifications for authorship. With researchers choosing to copy and paste materials without even removing the prompts that were used or the final closing comments at the end of the LLM output — can we say those authors have meaningfully taken responsibility for the final text? “I prompted a computer and maybe did some Make It Sound Human editing” doesn’t strike me as exceeding the sweat-of-the-brow argument for copyright. (There is a separately question about intellectual property and who it resides with if someone is using one of the open LLM models but I will leave that to the lawyers for today). The use of these tools has been shown to have higher risks of data falsification as the various algorithmic tools invent charts, fill in data that doesn’t exist, etc. Faculty will need to verify this consistently and document that added data hasn’t been introduced or fabricated. This is particularly critical as the tools are written to return a confident-sounding, even if entirely false or non-existent or changed result. We have some but inconsistent consequences for publishing false data and these may need to be increased as this is further damaging public trust in scholarship. For junior faculty, reliance on LLMs has the potential to demonstrate a lack of independence as a scholar. Use of these tools needs to be justified to address how you are overcoming this and why institutions should invest in researchers rather than solely having one senior faculty member just prompt and extrude some new text. If anyone with a vaguely decent prompt can generate your entire research agenda, why should the institution invest in expensive humans or award tenure? The belief in this is already floating around and is likely to lead to further de-professionalization and adjunctification. We need to provide documentation to external reviewers, potentially with annotations, about how and if generative tools were used. I’ve already started asking for this from institutions in external reviews and I think this will become common. Here is another instance where you will be asking me to spend a significant amount of my time and expertise. If the candidate couldn’t be bothered in the preparation of their dossier to meaningfully invest in this process, why should I? And while this is last and brief, it is no less important that as faculty, we need to consider the ethics of using tools that have enormous ecological impact with water, land, and power; requires extensive filtering labor from the global south; is showing significant risk for causing mental harm through sycophancy; and facilitates technoableism. A Few Suggested Materials Weaver, K. (2024). The Artificial Intelligence Disclosure (AID) Framework: An Introduction. College & Research Libraries News, 85(10), 407. doi:https://doi.org/10.5860/crln.85.10.407 Guest, O., Suarez, M., Müller, B., van Meerkerk, E., Oude Groote Beverborg, A., de Haan, R., Reyes Elizondo, A., Blokpoel, M., Scharfenberg, N., Kleinherenbrink, A., Camerino, I., Woensdregt, M., Monett, D., Brown, J., Avraamidou, L., Alenda-Demoutiez, J., Hermans, F., & van Rooij, I. (2025). Against the Uncritical Adoption of ‘AI’ Technologies in Academia. Zenodo. https://doi.org/10.5281/zenodo.17065099 The Data Workers Inquiry Project from DAIR https://dair-institute.org/projects/data-workers-inquiry/ Damian Williams “Virtual Power: Al, Ableism, Racism, and Fascism” at Philosophy, Disability, and Social Change Authorship and AI Tools. COPE. https://publicationethics.org/guidance/cope-position/authorship-and-ai-tools Bender E and Dair A. The AI Con. https://thecon.ai/ [This book is where I’m pulling the phrase extruded text from]