100 thoughts on “I am no longer chairing defenses or joining committees where students use generative AI for their writing”

1. I have a question and a comment.
   The question: What are your thoughts on AI-driven writing companions, such as DeepL Write? The technology is similar in that it uses large neural networks trained on huge corpuses of text, but it only edits text that has already been written. It does not alter the content.
   The comment: Last week, I received some valuable advice: ‘Write the first and last drafts yourself. For any drafts in between, you can use generative AI, but write the first and last drafts yourself.’ I feel like that would be a fair policy *if people would actually stick with it.*

   Reply ↓

   • Also, no white-on-white text with tiny font–unless it’s absolutely necessary because you really need to get that conference paper accepted.

     Reply ↓

   • The distinction to me is not whether the text was written and then edited versus completely generated by AI. It’s about how much editing happened. I have been reading MS Word documents for years where people added all the commas that word suggested and that has been okay but if you used generative AI to help you phrase sentences or re-order stuff and then ask me to review it as though is all your text, I don’t think I am okay with that and — as I understand it — DeepL can do just that. But correct me if I am wrong here.

     Relating somewhat to Andrew’s comment and something I was thinking while writing this, do we have (or not have and need) new terms for the different types of AI text? AI generated text to me includes: written by AI, heavily edited by AI (to the point of needing to be quoted according to me) and lightly edited by AI (to the point that non-AI products could have done it before AI) but we don’t have clear words to communicate these differences to one another.

     Reply ↓

     • Re: Writing Companions. A writing companion such as DeepL has two functions: it corrects grammar and spelling, and suggests alternative ways to phrase a sentence (e.g. by rearranging parts of the sentence, adding or removing words, or replacing phrases). I can then select the most suitable option.
       I pipe almost all of my important writing through DeepL, not because my grammar is terrible, but because I am a perfectionist. I want my text to convey exactly the message I intend. In my opinion, this is only a small step up from MS Word’s grammar check, which has become increasingly similar to DeepL Write in recent years. In DeepL, you can either accept the machine’s suggested improvements or put in the work and go through the text sentence by sentence, pondering the message and delivery of each sentence. I can see no downside whatsoever in using such tools, unlike generative AI, which can generate text without anyone putting thought into the writing process. Even if someone is lazy (or short on time) and just accepts what the machine produces without giving it much thought, at least the text will have correct grammar and typography. In such a case, the problem is not the use of technology, but the insufficient allocation of time.
       I believe this last point is really the crux of the matter: many people are trying to save time using these tools. However, saying that you cannot use such tools also means that users will not learn how to use them appropriately outside the classroom and does not strike at the great of the problem. While I can see the point of prohibiting generative AI to combat thoughtless writing, I feel that writing companions can have the opposite effect. I would encourage anyone to carve out an exception for them.

       Reply ↓

       • Raphael –

         I use AI in a way that’s similar to what you describe.

         I use it to more clearly, or precisely, or concisely convey what I’m trying to convey. In using Ai in this way, I often develop my thoughts more deeply. Often, I reject the AI revisions even if they might be more conventionally phrased, because those revisions lose certain nuances or style choices I consider important even if they decrease readibility.

         While I appreciate Lizzie’s concerns, I benefit from this process, and I with AI I can produce writing that better achieves my goals.

         That all said, I think there’s a problem whereby many students might lose out in the end if they become overly reliant on using AI in the writing process. I think it’s possible that while it helps them to produce a better-written product, it’s use will actually impoverish their learning processes, and in the end leave them less able to write and importantly, THINK, than their counterfactual non-AI using twin.

         It’s a conundrum. I like your idea of a structure where neither the first draft nor the final draft are substantively AI generated, although I think there’s inevitably a lot of ambiguity there.

2. Lizzie:

   Regarding your suggestion that “we write A LOT LESS,” I have some thoughts.

   If, when you say “writing,” you are referring to people spewing things out from a chatbot or copy-and-pasting from a textbook or wikipedia or whatever, or just running their mouth without thinking, I agree. There’s a place for such writing but it should not be published (and a Ph.D. thesis is a form of publication).

   But if, when you say “writing,” you are referring to the effort we put to clarify ideas, I don’t think we should do a lot less. Most of the work I do is writing, and I think it’s valuable. It’s not clear to me that I’d be making more useful contributions if I were to write less. And I’m not talking about the writing I do that serves purely to broadcast existing ideas; I’m talking about writing that’s essential to the development of new ideas or to the understanding of existing but perhaps nebulous ideas.

   Indeed, I think that most of my blogging serves this purpose too: it is by writing that I explore and organize my thoughts. Writing logical natural-language text is not so different from deductive mathematical reasoning: the point is to state my assumptions, work out their implications, and confront the places where these implications do not comport with reality or with my current understanding of the world.

   Reply ↓

   • There was an ‘if’ at the start of the sentence you’re quoting. The ‘if’ was meant to convey that we need to make some decisions here and I think that ‘if we want AI-generated text to be acceptable’ then we should change more than just saying it’s fine. And maybe we should change what is ‘writing,’ because when I read a manuscript or thesis I assume I am reading natural language. But I know I probably should not assume that, and part of the reason I should not assume that is journals, universities and advisors of thesis students are saying nothing about what is okay, or not, and raising no real ethical red flags about what we’re doing here.

     I think I agree with you. `Writing,’ as the word stands now, to me is a way to explore and organize ideas, to build and re-build them, and defend them. To me, the writing in a good dissertation thesis should do just that and represent a student’s own ability to do that. I don’t think we should do less of that, but I think we already are — and most of us have signed off on it (literally, in the case of doctoral theses).

     Reply ↓

     • > `Writing,’… to me is a way to explore and organize ideas, to build and re-build them, and defend them.

       I wonder if contemporary students feel this way about ‘prompting’?

       Reply ↓

3. predictions:
   – this rule will lead you to chair fewer defences and join fewer supervisory committees
   – this rule will preclude you from chairing any defences of supervising any students who are not native or fluent english speakers
   – if widely adopted, this rule would lead students to lie about whether they used GenAI to edit their writing

   Reply ↓

   • Short anecdote:
     We used to prohibit the use of generative AI at our chair. Nobody cared. Then, five out of eight students failed the seminar due to unauthorised use of AI tools, and a sixth student was not penalised due to the minor nature of the infraction. The problem was that the AI-generated texts contained hallucinations, non-existent sources and misattributed statements. Now, we are working on AI guidelines to educate our students on the appropriate use of AI tools. Our hope is to ‘get ahead of the wave’ and guide their use rather than suppress it. By ‘get ahead’, I actually mean ‘not get crushed again’😢

     Reply ↓

   • I don’t see you offering acceptable alternatives. There are people who will lie about plagiarism. That doesn’t mean we have to condone it.

     Reply ↓

     • plagiarism is much worse than using GenAI to help someone write, especially if they are not a native english speaker

       Reply ↓

   • I was worried about your first two points also. I don’t want to help non-native speakers less. However, in discussing this before I wrote the post, I found out that certain disciplines already do not allow generative AI for writing of theses at my university so it’s not that I will chair fewer defenses, it’s that I will chair fewer defenses in the the sciences — where we all appear to be ignoring this problem. And that reality is my problem: that we don’t value writing (natural language writing) enough in my field compared to others.

     Regarding your last point, I am not sure how much I want my rules to be determined by whether you can lie to me to overcome them. I am still — for better or worse — hoping for integrity and honesty from most of my academic colleagues, including all the students and trainees I work with.

     Reply ↓

     • in this case i think it would be good if people lied, because using GenAI to help you write is not intrinsically or necessarily bad. relying on GenAI too much–for example, having entire sections fully written by ChatGPT and which are not carefully checked for accuracy–is very bad. but relying on it a little bit–for example, as a way to correct spelling or grammar mistakes, offer feedback on style, help make writing more concise–is perfectly fine, and lying about this kind of acceptable use would be a victimless crime. in that sense, i think it is a bad rule, so lying to evade it is ok.

       english is my native language and mother tongue, but i work at an institution where the language of instruction (and the first language of most faculty) is not english. working in my second language (in which i’m roughly fluent) has made me much more understanding of the challenges non-native english speakers face in academia, and how valuable GenAI is in writing efficiently

       Reply ↓

4. Are advisors regularly editing texts for students for other reasons than being a co-author? I’m surprised! If it’s marked student projects, why would I edit them?

   I like your approach by the way; I had very similar thoughts recently. Of course there are problems with it, but there are problems without it, too.

   Reply ↓

   • I often offer to proof/edit texts for students when I have no intention to be a co-author. Because their projects are cool … and I want to support the student and their mentors (my colleagues.)

     Reply ↓

     • What I do is I send them feedback on writing, occasionally pointing out specific errors. Of course it’s nice of you to help, particularly if the thing is later published with you as a co-editor or not, but (a) it’s certainly not part of our job, and (b) if it’s a marked project and I give too much help, it’s myself who is marked rather than the student (of course it’s a fine line because project feedback is an important part of student learning, but editing goes too far in my view).

       Reply ↓

       • I realise that my comments were written with more general student projects in mind whereas the opening post seems to be specifically about PhD.

5. I largely agree with the sentiment of this post, and echoing some of Andrew’s thoughts, it feels like using GenAI to edit and/or make large swaths of text takes a lot of the understanding of how to logically and clearly present challenging material out of the picture. That seems bad, especially because (as I understand) the thesis has (roughly) two prongs: 1) show us what you did, 2) explain clearly the subject of study, problem, and what contributions you’ve taken towards solving the problem.

   Further, I’ve seen what over-reliance on GenAI for writing does to people – at my job, I have had projects where the boss will meticulously input your writing, presentation text, etc. into the corporate chatbot and ask the chatbot to “summarize this text”, resulting in my frustration when pieces of information I thought were important are left out. When I ask this particular person “Why did you take X out?”, I frequently get responses such as “Well, it didn’t read well”, “It was just kinda long…”, etc., but no concrete advice such as “This audience is less technical, they won’t need additional context” or “We don’t want to elaborate on an experiment yet because it hasn’t finished running”, or anything with such reasoning. So I can see how using this technology in a thesis could make people think much less about why they’re including or excluding information, why certain sections are summaries whereas others need further elaboration, etc.

   I also agree that it seems silly to edit largely GenAI created text, because at most you’re affecting the student’s future prompts, not writing. And if the reviewer is using AI tools to edit too, what a silly little roundabout the whole process of editing becomes.

   However, I don’t necessarily agree with some points. First, the response to the point “Asking people to quote from chatGPT or similar is insane, no one will do that.”. I don’t think that quoting/not quoting from a tool that is intended to help edit papers is necessary. For example, would you have a student quote a single additional word in a sentence, and reference ChatGPT? That feels like making someone quote Microsoft Word’s grammar checker for helping with the addition of a word, or adding “-s”, “-ing”, or “-ed” suffixes for words. I understand many people use these GenAI models to make large swaths of text, and this is a distinct example from my point. In my opinion, when a chatbot helps write the paper in a substantive manner, it should be a conversation between advisor+advisee to determine if the use is appropriate, and then mentioned in the publication.

   Second, and maybe I’m nitpicking here, but the quote “Not to mention, all that text we’re working on is thus being rapidly fed into the models owned by private companies.” caught my eye. Isn’t it true that this will occur regardless of the paper’s origin? If a student+advisor write it, it will end up in a dataset somewhere. If a student+advisor use ChatGPT to write it, won’t it end up in the same dataset?

   At any rate, it’s a weird time to be in academia, and I appreciate you putting this out there for a proper discussion to start (hopefully) – and I generally agree with you!

   Reply ↓

6. +1

   It could be the case that the quality of output from Chatbot + Human > Human. If so, then the question is, why should we not use it? There could be pedagogical reasons (concern that students are not learning enough when producing the output) or it could be that students are abusing the system in that they produce a lot of volume with variable quality even though the mean is high, and judging quality is hard, and this leads to burden. Or it could be that, based on data, most student outputs using bots < student alone, as students are poor evaluators of bot outputs and assess in a cursory manner.

   I think the issue you raise is the last one. My hunch is that the policy space is larger than honor policies. For instance, if the issue is over production, you could do it in the classic expensive talk way that you will only give feedback on a set number of things each time period and it is up to the student to pick those essays. And that there would be a 'negative marking' system for all bad points?

   Reply ↓

7. In the future, producing several hundred pages of prose will no longer count as a performance of any relevant academic competence. Instead, we will have to require a brief synopsis of the results the student is prepared to defend. Then we can design a series of written and oral examinations to test them (in a proctored, off-line setting). I would argue that writing those several hundred pages (without AI assistance) will remain the best preparation for such exams, but actually inspecting those pages will tell us nothing (because we can’t tell what AI contributed).

   Reply ↓

   • Strongly Agree with Thomas Basboll. Allowing use of AI reduces competence of the student who fails to write based on their own knowledge and understanding. This is such an obvious consequence that it renders many of the other posts on this issue to be very faulty guidance.

     Reply ↓

     • It’s a catch-22 for me. AI has the ability to act as a force multiplier for someone who is already good at what they want AI to do to augment them. But in order to become someone with that capability, one cannot use AI to substitute for their learning and experiential process to get to that point. In education, I believe this means as you learn, you want to be able to connect the dots yourself but have AI take care of the things you can now do in your sleep since you’ve mastered it. In some ways it’s like teaching or mentoring. As a teacher, I want to help you learn by introducing obstacles and providing helpful suggestions at timely moments so that you can grasp the full context and ramifications that I had to think about to give you those timely bits of guidance. Today we are so fearful of the future ramifications that we aren’t fully embracing the headaches and hardship now.

       Here’s a thought experiment. Let’s say AI gets to a point where it can write better dissertations than any human can as far as following and passing a standard rubric for excellence is concerned. The only thing that matters at that point is in the specific novel insight or new-ness of a thought that never existed before, not the prose, not the research, not the clarity or conciseness. AI would handle all of that. The purpose for which people write dissertations is what exactly, in this future?

       Reply ↓

       • Sang:

         I take your point, but I disagree with your statement regarding the dissertation that “The only thing that matters at that point is in the specific novel insight or new-ness of a thought that never existed before, not the prose, not the research, not the clarity or conciseness.”

         The main point of the dissertation is not its content; it’s that it prepares students for independent research in the future.

         Again I recommend the analogy with textbooks.

         Suppose the thesis were to contain an explanation of some method. The point of asking the student to write this explanation is not to make the thesis a wonderful and well-written document; it’s to ensure that the student understands the method at hand. This writeup serves two functions: first, the act of writing something up in your own words is a way to understand a method; second, for the evaluators it demonstrates that you understand the method to the extent that you could write it in your own words.

         An alternative would be to copy and past material from an existing textbook (with proper attribution, of course; otherwise it would be plagiarism and would be grounds for kicking the student out of the Ph.D. program entirely). Copying-and-pasting from a textbook (with attribution) could satisfy the goal of making the dissertation more understandable to the reader, and that’s not nothing–but it does not represent any steps toward the much more important goal ensuring that the student understands the material.

         So, as I see it, the purpose of a dissertation in the future is the same as the purpose of a dissertation in the past and present, which is to (a) prepare a doctoral student for a working lifetime of doing independent scholarly research, and (b) demonstrate to the evaluation committee that the student is indeed ready to go out in the world.

         Chatbots, AI, etc., can be useful for these purposes–not by producing prose which can be inserted in to make the dissertation more readable, but as tools to help students with their learning, in the same way that textbooks serve is goals.

   • Interesting to me that there are lots of opinions on the details of what should be allowed for certification, but there are zero people saying that the role of certification is basically a bogus concept.

     Far far too many people buy into the idea that a major important role of universities is to stamp approval onto job candidates so that the people in charge of companies will know who to allow to do high paid jobs. Hierarchies ruin everything. Ive got a PhD student in Germany, he and I collaborate on a research project that hes using to get a certification (ie. a PhD) but I play no role in the certification. I’m just teaching him how to build models, fit them, and present them. The project itself will be evidence enough of his abilities. The university is a complete superfluousness. Of course everyone’s gotta convince some “boss” to pay them so they can eat. But ultimately, thats the problem. Noone would be cranking out AI generated bullshit if we weren’t all begging for permission to eat from a boss. Thats why i consider myself an Anarchist. I have no use for the “bosses” who turn everything into a meta game of how to game the system to convince someone to let you live.

     Reply ↓

     • I’m fond of anarchists, and I wish them well most of the time. But I’ve not yet given up on the university and the “academic” way of knowing that it maintains: in a nutshell, the sort of thing you learn by taking a class and reading books, an understanding framed by “what is known” (by the culture).

       You can’t have an institution with “degrees” and “chairs” etc. without some form of assessment. The problem now is that traditional indications of distinctly academic competence (which is essentially discursive), namely, reading and writing a particular kind of prose, can be generated without that competence. That is, a scholar’s prose can no longer be presumed to represent their knowledge. Unless, that is, we can be certain that it was produced by the scholar’s own mind and body in a space cut-off from AI.

       Reply ↓

       • Dont get me wrong, I agree you can learn a lot from a class etc. I just dont think that universities (at least here in the US) primarily trade on their ability to teach those classes. I mean witness the rise of giving away almost all of the class content online…videos of lectures, lecture notes, etc the MIT open courseware and open textbooks and Kahn Academy lectures and etc etc are all free to access… What people pay for is the stamp. I dont think thats a legitimate business myself and in the long run I think it’ll collapse when people realize the stamp has increasingly become worthless.

       • It’s true that many people approach assessment as the basis of a “stamp”, a signal. But there’s also a more authentic reason to think about assessment: the sincere student wants to know whether they are any good at the sort of research they are learning, including whether they are good at putting it into words. So there has to be a way of demonstrating actual competence/knowledge to someone who already has it.

         With AI in the loop this is almost impossible. When the output is correct, we don’t know what it indicates about the writer. When the output is wrong, we don’t know how to begin correcting them.

         Andrew compares it to saying only what Wikipedia says. This is apt because, although Wikipedia is right about many subjects, we can’t help a student learn if they only tell us word-for-word what they’ve found there. They have to show us what they think is true and how they think it is best expressed in sentences and paragraphs.

         That is, AI output doesn’t provide a good basis for the sort of learning you get by correcting errors.

       • I agree with you. A student who actually *wants* to learn, will not use AI tools to generate responses to queries designed to help them determine if they understand. A student who wants a stamp will do whatever it takes to get that stamp. When the stamp determines how much you can eat, have children, or go on vacations, the stamp will be the primary thing people are motivated by. Only by changing society will we fix the illness of people learning little but getting stamps.

         In the absence of that illness, the problem doesn’t arise nearly as much.

         Much of society is organized around this principle: assets are under control of small elite. Elite has limited information about good uses of those assets. Elite seek signals to allow them primarily to increase their assets through use of labor resources. Laborers send mixed true/false signals to Elite to gain access to income needed to live / enjoy life. Elite make a number of fairly random choices, get fairly random outcomes, blame bad luck when things go wrong, take credit when they succeed.

         the general public learns that material conditions depend on signals, luck, and connections, and to a lesser extent skill, and respond accordingly. The signal generators learn that they make money primarily in proportion to them taking money to give signals… the randomness / quality of the signal isnt that important because the consumer of the signals (elite and their managers) aren’t super clear on the value of the signals anyway… it’s all basically comes down to the wrong people (elite / managers) are making decisions with very poor information.

       • +1

         Daniel gets it (re: stamps and asymmetric views of luck).

         However, I would argue that even a student that wants to learn very well may use AI to generate responses to genuine pedagogical queries / assignments. The difference is that they would interrogate the output until they were satisfied, and under the assumption of a caring / diligent student, that is not a low bar in the least.

       • Replying to Daniel: My only solution to the general problem you’re (more or less rightly) identifying is something like Henry George + C.H. Douglas (i.e., land value taxation + universal basic income) but I think that’s the long way around the immediate problem that AI presents us with in higher education, which can be solved (in my opinion) simply by testing students offline.

         Replying to AllanC: I think I’m with Lizzie on this. If a student wants to learn anything from ME, the deal is going to have to be that they show me the words they’ve put together themselves with their bare hands (and brains). If they want to talk to an AI too, that’s fine; just don’t ask me to read it. Don’t even quote it at me. Tell me in your own words what you think and I will tell you in mine what I think of that.

8. This is a fascinating topic. Software development was the obvious field in my mind where LLMs have impact, now it seems that academic writing is among the early casualties as well.

   Or writing in general. Beyond information delivery, it has accumulated many other functions, including thinking (like Andrew said, and I agree), proof of cognitive mastery of a field (exams, essays), a collection of quality and commitment signals for evaluation (various applications), or just proof of use of time, showing commitment, maybe other less obvious ones. LLMs disturb these arrangements and implicit contracts involved (“you write, I edit for your benefit”), either by producing outright fakes with surface-level similarity only (so called AI slop) or producing quality practically indistinguishable from humans.

   And I have understood that really distinguishing AI text from human text is practically impossible at the college level already, especially given that there are public services for AI detection through which one can run the text before sending, and also humanizer services that add typos and other non-AI characteristics. (Currently the difference is probably more obvious on higher academic levels.)

   An example wider than academy are e-mails or reports that no-one writes (beyond a few bullet points) and no-one reads either (beyond AI summaries). It’s easy to tell that producing such documents is not the final equilibrium state of the system.

   It may take a while until all this is sorted out.

   (I’m a statistician or ML/AI expert in the industry; I have however proof-read so many scientific papers that I can symphatise with both sides: those you see AI edits as a relief from low-level editing, and those who see editing AI text as a waste of time).

   Reply ↓

9. Your rules for what is and isn’t plagiarism seem quite arbitrary. Appeal to “guidelines” is not a defensible reason for moral indignation by itself. You say that you “feel” that such and such use of artificial intelligence is alright, but that such-and-such is not. I suspect that your graduate students chafe under your vibes-based supervision. I know I would. If PhD candidates in the humanities are professional-level scholars, why treat them like wayward undergraduates and forbid them from using a powerful new research tool? Why draw the line at “grammar checks” and forbid perfectly legitimate use cases like “Would this sentence read awkwardly to a native speaker? Should I move this clause here? What do you think and why?” On the other hand, if your graduate students are tempted to abuse that tool to avoid the process of thinking and writing for themselves, why are they in your program in the first place?

   Reply ↓

10. You do know those was this written by AI tools are incredibly inaccurate and give a ton of false positives…

    Reply ↓

11. I was surprised to see that you’d be okay with AI generation of code. I feel that code describes the methods that are actually applied, much more than a description of what was intended to be done. I wouldn’t want someone to cheat/plagiarize a description (use AI in writing) but I definitely wouldn’t want them to not have a detailed understanding of all of the nuances of the method that is actually applied (the code).

    Linters, debuggers, static code analysis, and the like should be capable of helping with the syntax while leaving the programmer to do the real work and use real brain cells.

    Was a previous generation’s similar conundrum the use of calculators in math class? Learn to do the division and multiplication by hand … only by doing it will you understand what the steps are and can understand why they’re done. Of course, now I often reach for a calculator when I have to do math; but I do try to show my work if it matters (if only for my later reference.)

    Reply ↓

    • There are different kinds of code. “The method” is often a minuscule part of the whole, among a broad class of “I/O” (APIs, DBs, data formats, graphics) and skeleton (think of all the code needed to run a Stan model, which now easily comes from an AI, both on the R side and on the Stan side).

      While writing the core manually is defendable, I’m not sure that applies to the rest.

      And there is no clear line between AI and the coder. In practice it’s a complex piece-by-piece iteration.

      Reply ↓

    • I didn’t write that I am okay with AI generation of code. I wrote, “These guidelines apply to your writing of text, not your writing of code.” Code is a different ball of wax to me. And I am not asked to read the lines the code underlying a PhD thesis often, I am asked to assess the output.

      Reply ↓

12. This is really interesting. I’ve thought about this at length too, because I personally find it hard to imagine thinking without writing. Part of me wishes I were able able to insist on similar guidelines. But reflecting on what stops me, I think it’s that regardless of how it was written, the student preparing the work is ultimately responsible for how much sense the written text they submit makes. If they can’t see that they are laying it out in a subpar way or being sloppy and vague with language, then it doesn’t really matter whether it was their original attempt or a mishmash of them and generative AI, they still deserve an opportunity to have that pointed out. Assuming I’m choosing my students well, either way I expect the final product to improve in light of the feedback they are getting. In other words, ultimately it’s up to them how reliant they want to be on an AI-assisted workflow, and I will just continue to expect text that meets my standards, because that’s what I care about.

    I say all that with mentoring PhD students in mind. GenAI use can also be frustrating in teaching. When there are assignments that involve writing and the bar can’t be as high as I might expect in writing research papers, then I can get annoyed by the possibility of handing out passing grades to students who never even tried to write themselves. But again, I’m more comfortable leaving them responsible for their own learning than trying to control it. If they want to try to become great academics by never actually writing for themselves, I say let them.

    Reply ↓

    • Jessica –

      Your final sentences remind me of when I was working at a university where there was a lot of attention being focused on students cheating on tests. All of the focus was on how to prevent students from getting away with it. None of the focus was on what it said that many students felt that cheating on tests helped them to achieve their goals.

      Reply ↓

      • Joshua:

        For a lot of students, their goals would be achieved by just being given the Ph.D. without having to do any work. Then they’d be qualified on paper for any job that requires a Ph.D.

        This sounds kinda silly, but it seems to have been the career trajectory of various people such as the former Dean of Engineering at the University of Nevada.

        Getting a Ph.D. served this guy’s purposes but I don’t think it served society’s purposes, nor did it serve the purposes of the taxpayers who funded the universities where he worked.

        My point is that individual students’ goals are part of the story. But only part. Setting aside all issues of honesty, rules, and cheating, for any institution we need to think of larger goals, not just the immediate goals of various players.

        Reply ↓

        • If lots of students just want the PhD without work, that’s a problem with the system, no? We must be failing in communicating what a PhD is about. Ultimately if those kinds of students are getting PhDs, with or without using LLMs to write, that seems like an issue with the advisor not paying enough attention to have a sense of what their students are capable of. Eliminating LLM use won’t solve the real problem, although maybe it makes it easier to spot these students early (because they won’t try very hard to write themselves when required). But that supposes that you can reliably detect when they’ve used an LLM to make sure they are complying, which requires a huge amount of effort, and is probably still a losing battle.

        • Jessica –

          My view is similar.

          At some level, the whole dissertation process has a similar tension. On the one hand it’s a process of learning for students – to create a unique product that requires a mastery of the literature and a deep level of analysis. On the other hand, it’s a hoop to jump through -where often the product is not particularly meaningful beyond the defense (how many dissertations kienundrrad after the student has graduated – that exists largely because their professors had to jump through the same hoop.

          I don’t mean to trash the entire system – but often we focus on the trees and not the forest. Using AI can support students’ learning, but only to the extent that matches the intent of the students. Taking away the Ai wouldn’t really address the underlying importance.

      • I would second this view. Especially since other hurdles early on in the process (e.g. comprehensive exams) ought to be able to ferret out those that use AI as a replacement, rather than a supplement, to their own learning. If that’s not the case, there is a much bigger problem with the system. In general, I don’t much care about how someone conjured up work product, as long as it’s of a high standard – save for stealing and the like, which to be fair, I know is a contentious issue with AI. The biggest hurdle with employees (or in the OP’s case Ph.D. students), in my experience, is that people don’t even know what good product is supposed to look like; and that is not for lack of being shown exemplary examples. And at this moment, AI is nowhere close to being able to produce good work product in the domains I am familiar with without a human filter overlaid over the output. Though, of course, that may change fairly rapidly.

        The above said, Lizzie does have a good point about not wanting to waste her time reading drivel. But I would submit that students who produce drivel using AI would have done so without it anyways. To me the crux of the problem lies with the admissions stage. Admit students that want to learn, and AI use should be of no-concern as they’ll use it responsibly – as a learning supplement. Admit students that don’t have a desire to learn, and they’ll use AI irresponsibly, but even if they don’t use AI, they’ll find otherwise to cheat, lie, and steal their way through the program (hence other hurdles required – see above).

        As an aside, your comment about thinking while writing is reminiscent of the divide between structural engineers and drafters (I’m a structural engineer). Large firms are generally set up so that drafting and engineering are separate departments ostensibly to take advantage of division of labour. However, just like you do much of your thinking while writing, I do considerably amounts of my engineering while drafting. I cannot possibly separate those two functions (I have tried and often ends up in more work in redlining the drafter’s output anyways).

        Reply ↓

        • Interesting about engineering and drafting. Maybe the bigger tension, which writing and thinking and drafting and engineering are exampes of, can be described as design vs implementation, and the lesson is that ultimately this become hard to separate in lots of highly skilled work. So having a good simulator for one will never really be enough.

13. Three questions:

    1. What about generative Ai when writing code? (If not, how is code different?)

    2. Absent verification, do you worry that you will select for liars?

    3. Given scope for lying, should we include oral and written examinations after the thesis has been submitted.

    My answers:
    1. No, but coding is journey.
    2. Yes
    3. Absolutely, otherwise we will certify nit-wits.

    Reply ↓

    • If I understand you, I agree on AI coding.

      That is: if the student debugs the code and demonstates (through sensibly thought out tests) that it does what’s needed and/or claimed, then AI assisted coding is fine. The student has fully solved the problem given, and has taken responsibility for the work presented doing what’s claimed. I’d argue that the student should have to defend the test cases as part of the thesis defense, though.

      The selecting for liars bit is a brilliant thought. And your solution sounds excellent as well, albeit only applicable to the PhD program context.

      Reply ↓

      • Thanks David. I appreciate the kind words. I agree with you on coding.

        For context, I’m only teaching graduate students in New Zealand, but this covers the Honours degree (equivalent to year 4 in the USA) through to the PhD.

        For Honours students;
        – I’ve reintroduced in-class tests (remember those?) and
        – whiteboard presentations.
        – I have no declarations or bans on AI for final reports because my sacred duty is to prevent cheating, not reward it. Cheaters will cheat, honest kids will declare. I mark harder.

        For PhD students:
        – If we want the degree to carry a signal, we need in-class examinations as part of the thesis defence. I bet writing such questions would tempt many of us, perhaps even Andrew, into assessing PhD theses again.

        Reply ↓

        • Jb:

          The thesis defense at Columbia is in person. So I’m not quite sure what you mean by your last sentence.

14. Hard situation. I understand your frustration but just a different perspective. I am not being sarcastic but these are not different situations: Can they use calculator? Or maybe you should ban them to use computers and internet as well? Maybe they shouldn’t use NumPy and do the linear transformations for SVD by hand? Or should they submit their thesis in hand-written form, no printers? Bottom line, If there is a technology at almost very low cost, people will use it. The genie is out of the bottle.

    Reply ↓

    • MS:

      I think the difference here is that you can be almost certain a functioning calculator will give you the correct answer. Malfunctions can occur of course, but they seem exceedingly rare. AI, while remarkable, are still very error-prone. Whenever someone uses AI in their work you have to trust that they carefully verified the output, and that’s not really a safe assumption in my cynical opinion.

      Also, there’s a difference between a purely mechanical tasks (like linear algebra, or printing a page, etc) and more creative/individualistic/expressive tasks like writing. A lot of people value that. I personally don’t, at least not in technical manuscripts, which is why I still use AI assistance for editing and coding (not much honestly, but still more than Lizzie would find acceptable I think, especially given I’m a native English speaker).

      There’s also the fact that AI is sort of annoying. It’s being shoved everywhere. Even as a pro-AI leaning person, I’m getting sick of it myself.

      All that said, I don’t think that AI assistance, on its own, should be considered reasonable grounds for refusing to engage with student training (although I do agree that AI should be credited when used). I think it’s surprising that Lizzie specifically excluded AI assistance for code generation in her guidelines; I would personally consider that a greater risk to technical correctness than AI assisted editing. I’m curious if Lizzie (or anyone who shares her views) uses AI for her code, and her justification for it given her distaste for AI assisted editing.

      I was kind of tempted to use AI to edit this comment (with acknowledgement) as a joke, but that might have come across as a tad provocative.

      Reply ↓

      • 9P:

        Great analysis. Thanks. Yes, reliability aspect is another dimension of this larger problem of using GenAI in academics.

        > “…there’s a difference between a purely mechanical tasks..”

        That’s a bit metaphysical. I value this approach a lot, that human creativity is distinct but the line between mechanical task and human creativity is now blurred due to LLMs at some degree, at least compare to pre-LLM era. There is still a human design aspect of implementing mechanical tasks: such as implementing numerical algorithms and their data structures, i.e., NumPy example. Just trying to see from other aspects: if we use automation in other parts of the research process why not in writing, at least as a grammar checker and detecting fluency in writing and recommending more explanations where we missed.

        Reply ↓

15. Frustrating. Admired the principled attitude. The technology of detecting AI generated text vs. human written is also advancing. Strong guidelines and onboarding graduate students on ethical work and integrity maybe helpful.

    Reply ↓

16. Hi Lizzie,

    I actually found a use-case recently where I recommended to the authors that they have an LLM fix their English. The paper was submitted to a journal and I was a reviewer. The paper was written by non-native speakers who have no access to a native speaker advisor/co-author, and it was full of incorrect use of determiners and other grammatical errors that an LLM would easily be able to correct. The work was well executed, but not well written. I don’t see the problem in allowing non-native speakers to clean up their writing using such tools (but maybe you would agree with me on this kind of usage). I guess it’s hard to prove that their first version was written by them without any aids. But one has to trust the authors to do the right thing…

    PS I use DeepL to translate text into German, then have it corrected by a native speaker, when I have to write formal reports in German.

    Reply ↓

    • I was going to ask if the journal has an editor, and if so why not let them correct the language. Your PS suggests that this is what should happen, with the editor possibly using LLM to speed up the job. The downside for you, the authors in your anecdote, and the editor (assuming they aren’t super-fluent in the technical details of the work), is that none of you could read your text and be sure that nothing was lost in translation.

      Reply ↓

      • Apols for butting in but it’s an interesting question.

        The job of the editor isn’t to edit manuscripts they’re overseeing although they may well recommend that the authors seek editorial help. Some journals have in-house editing facilities but it has to be said that many journals seem happy to publish “as is” any manuscript that has been given the green light by the reviewers, and so some papers end up being published with appalling language/grammar. This is a shame for the authors and in these cases IMO someone is being remiss – either the editor/reviewers for not strongly recommending/insisting on some editing or perhaps everyone (including the authors) for not caring. As a reviewer I usually make some suggested edits for authors to consider.

        In the old days, when scientific paper submission was done by mail with hard copy manuscripts (“please include the original manuscript and three copies with your submission”), we used to get (especially from American Chemical Society Journals) the proofs with the subeditor-corrected manuscript back in the mail, and invariably would have to spend quite a bit of time recorrecting edits that a subeditor had made so that the original meaning was restored. That was always fun!

        Reply ↓

        • I certainly didn’t think The Editor (so to speak) of a journal would be running spellcheck, but I thought journals would have someone checking the grammar and so on when they lay out the article for publication. It’s been a while since I published in a journal but I had the same experience you did, so someone definitely made non-content-related edits.

17. I also heard an Odd Lots podcast in which they were interviewing a lawyer who enthusiastically uses AI (o3 or something like that) for preparing drafts, and was saying that o3 is really good at the job. But in the end he would read and edit using his expert knowledge. So the final product is a result of him weighing in with his expertise. This is different from what many people do when they use AI tools: they don’t have the expertise to judge the end-result. I think that if the user is an expert and can judge the accuracy of the AI-produced text, and correct/edit it where needed, that could be defensible. For PhD students, who by definition are still on the road to becoming experts, it could be hard (but not impossible) to get them to post-process AI generated text and still produce a thesis that counts as their work. It would only work if they had put in the effort to really understand the literature themselves.

    Among my masters students, those who are using AI to submit reports, some are not even reading the end-result of the report and have not read the associated paper either, which has the effect that some reports are mentioning experiments or claims that were not even in the papers being summarized or reviewed. For masters students, we are planning on introducing oral exams. I think this approach to examining students will be an improvement over written reports.

    Reply ↓

    • One thing I was thinking about was: those students mindlessly using AI, they will eventually crash and burn in the sense that they will only get so far after graduating. It’s an interesting natural experiment to see whether someone can become a tenured professor or a well-paid industry employee by just querying AI tools to get their work done for them. My guess is no.

      So in a way, this subset of people is wasting their time doing a degree. They could just do something else that doesn’t require them to learn about all that complex stuff.

      Reply ↓

    • Using AI tools, understanding the results and integrating them to the end product is exactly what students should be doing, in my opinion. That’s how calculators and other tools are used as well, and it builds the right skills, assuming the current technological level of AI. (That assumption is likely false in future; I have no idea how to prepare for that though.)

      Reply ↓

18. It’s hard to see how using AI to write code is okay, but using it to write text is not.

    I tell my students they can use AI in any way they want. The tools are advancing, the world is changing.

    Reply ↓

    • Carl:

      The question is, what’s it all for? Textbooks can be great (I say this as a textbook writer, but I really believe it); that doesn’t mean it would be appropriate for a student to dump an already-written textbook into a Ph.D. thesis. What would be appropriate would be to cite the textbook and quote as needed. Similarly, I find wikipedia very useful; any student who wants to use wikipedia in a Ph.D. thesis can read it as background or quote from it directly. It would not be helpful to do a Wegman and copy-and-paste without citation. If a student wants to use a chatbot to help work out some analysis or summarize some literature as part of a Ph.D. thesis, fine; what’s not appropriate is to copy in that chatbot output without attribution. When I say such behavior is inappropriate, I mean it in two ways: first, as Lizzie says, it violates the rules against plagiarism; second, it adds no useful content to the thesis, in the same way that copying a chapter from an existing textbook would add no content.

      Reply ↓

19. Rather than banning the bot use, it sounds like you want some kind of version control for these documents. Where each “commit” is associated with the bot, spellchecker, or a human (student, teacher, etc).

    Then you can skip editing any sections written by a bot. And part of the grade is now having good version control hygiene.

    Reply ↓

    • I think I would feel better if I could see all the edits. And I wish everyone did version control also.

      Reply ↓

      • Yea! The great thing about that solution is that this tech is already widely used, just not being applied to academic documents.

        You can still go back and read versions of wikipedia pages from a decade ago. Every document should be like this.

        Reply ↓

20. Janne,

    I entirely agree — embrace technology. Indeed, I am a power user of Claude Code. It’s the best technology since my Commodore computer in the 80s. However, to cultivate those parts of the educational process that require internalising knowledge and building skills, do you think we should introduce paper-and-pencil tests for PhD candidates *after* submission? In my corner of the world, the quality of PhD oral defense performances has dropped recently. I think it’s too easy to fudge them.

    jb

    Reply ↓

    • I don’t know. How to sort out various functions of writing, incl. performance evaluations, from each other in this new situation is a really hard question. I’m not teaching either, so not thinking this actively. If all non-monitored writing is in doubt, that does seem to suggest some sort of monitored tests or interviews though.

      In Finland oral defenses are mostly ritualistic and don’t serve as performance evaluation nowadays. Neither do I as a pretty stage-anxious person like them (although my own defense went fine). I have understood they are less public elsewhere and more about mastery of the field.

      Like I said elsewhere I definitely understand Lizzie’s point. I wouldn’t want to comment hundreds of pages of AI text either to supposedly help the student who is not there. Meanwhile, as a non-native English writer I have proof-read lots of scientific papers and theses which have been painfully bad writing on every level, so I understand those who say AI is a relief. Personally I use LLMs a lot for coding, “web research” etc. but very little for writing, as most of my writing is not that critical, and AI edits just seem to destroy my message and style. (Haven’t tried the latest models, nor specialized services. Occasionally I ask the models to spot worst errors etc.)

      Reply ↓

21. I think there are 3 somewhat different issues involved in this post and the comments:
    1. Should students (at varies levels) be permitted to use AI (and in what ways)?
    2. Should an instructor refuse to work with students that use such tools?
    3. How do we evaluate student work when AI tools are so available?

    Personally, I think the answer to 1 should be yes. It is another resource and I can’t see how we can distinguish between acceptable and unacceptable resources. Of course, I would consider submitting AI work directly (or with very slight modifications) as plagiarism, but using AI to edit work and then making decisions about what edits to use should be something students are permitted to do. I also don’t believe we should restrict who students consult with regarding their work (except where assigned work specifically says they should not use any resources other than their brain in a timed exam – though I see little value in such assigned work).

    Regarding #2, I think instructors should be permitted to set the ground rules regarding how they are willing to work with students. If Lizzie finds her time filled up with having to wade through AI generated text, then she has every right to say she won’t work with students that work that way. I have not found that situation myself, but everybody work in their own conditions and students vary in many respects.

    #3 is one that seems to occupy a lot of attention – too much in my mind. I didn’t go into teaching to spend my energy policing whether students are cheating. At the same time, cheating is real and can’t be ignored. My solution is to provide assignments that minimize the ability to cheat in a number of ways. I don’t think it is difficult to design assignments where bad use of AI can be easily detected – and I think those assignments are better than the ones that enable students to easily use AI in the worst possible ways. If we are talking about dissertations, then perhaps we just need to take the oral defenses (aren’t these still required?) more seriously. A student that has AI write their dissertation and ends up with fictitious references should find it difficult to fake their oral defense. Nobody wants the oral defense to become inundated with AI produced dissertations and the unpleasantness of flunking people at that stage of their work – but I do believe that making the ground rules clear can avoid such a situation. I would add that many such problems arise with scale: if you have 5 students to advise, these problems are more tractable than if you have 50. I would argue that an institution that asks instructors to teach a normal course load and supervise 50 students as well, is demonstrating a lack of care about quality.

    I think that failing to distinguish between these 3 questions stands in the way of clarifying how to deal with AI in student work.

    Reply ↓

22. I have no problem with a no-AI policy as a statement of principle, even knowing that it’s going to be hard to enforce. You gotta do what feels right to you.

    If the goal is to make the best use of one’s time as a reviewer, is there a threshold above which it’s acceptable to not engage in detailed editing? It’s not like bloated prose and sloppy organization didn’t exist before ChatGPT. I wonder whether there are situations where the feedback could just be something like “I can’t make sense of this; you need to cut the word count in half,” not because the draft was written by a bot (maybe it was, maybe not) but because it’s poor writing that places too much responsibility on the reviewer to try to pick out what’s important, like a traveler holding out a handful of unfamiliar currency and counting on a shopkeeper to take what is owed. My impression is there are still resources on campuses to help with general writing skills; is it practical to rely on that layer of support to help a writer cut through what may or may not be AI slop?

    Reply ↓

23. You’re essentially announcing that you’d rather spend hours editing grammatically tortured prose than minutes reviewing AI-corrected writing, all while claiming this saves time and helps students learn. Your policy treats spell-check and paragraph generation as equivalent crimes, demands Byzantine documentation for comma corrections, and somehow concludes that making non-native speakers struggle with grammar instead of focusing on their ideas is pedagogically sound. The real kicker? You’ve created more administrative overhead than the “problem” you’re solving, turned a functioning oversight system into evidence of crisis, and confused your personal preference for editing bad writing with principled academic standards. This isn’t rigorous mentorship; it’s performative nostalgia that actively harms the students you claim to be protecting.​​​​​​​​​​​​​​​​

    Reply ↓

    • Lzo:

      I think you’re misunderstanding the situation. The problem isn’t students writing “grammatically tortured prose” which is then fixed by a computer program that provides “comma cirrections.” The problem is students handing in a thesis with many pages full of essentially undigested chatbot output. It doesn’t matter if that chatbot output is grammatical or whatever. The problem is that it adds zero content to the thesis, or, more precisely, no more content than would be added by copying in a chapter from a textbook or copying several wikipedia articles.

      Reply ↓

    • AI can be used to fix grammatical errors, catch repetition, suggest re-ordering or re-organization of a paragraph or chapter, and other stuff that I think is great. I have a friend, formerly a magazine editor, who is fond of saying “everybody needs an editor”, and AI can provide a good editor or at least a pretty good editor. I use AI this way myself sometimes and would have no problem with a student doing so either, if I were evaluating a student’s work. I fall short of LZO’s suggestion that it’s shameful not to allow students to use the technology this way, but I would agree that I see a complete ban as serving no useful purpose in a lot of cases.

      Andrew says “the problem is students handing in a thesis with many pages that [contain] no more content than would be added by copying in a chapter from a textbook”, and I can see that that would be a problem, but I think that rather than completely banning the use of AI, a more useful approach would be insisting that students not include pages that add no important content, whether such content is included by the newfangled way of using AI or the oldfangled way of simply paraphrasing (or indeed plagiarizing) from a textbook. Banning AI because it _can_ be used badly is throwing out the baby with the bathwater.

      Reply ↓

24. Get of my lawn artificial intelligence because I think you are fake
    You are just what Cypher in the Matrix ate, which was a …..
    Are you capable to produce, complete, and/or understand this previous rhyme?
    And might this be something that shows why a human brain is sublime?

    Maybe it is the case that thinking and writing by a human brain
    Might be more unique, and complex, than what the mere words contain
    Communicating what’s not written might sometimes be what’s to attain
    And to do, and understand, this might particularly take a human brain

    Reply ↓

25. Howdy. I wrote my response here- https://mguhlin.org/2025/07/20/i-wont-help-if-you-used-genai/

    Thank you for exploring this topic.

    Miguel Guhlin

    Reply ↓

26. Just noticed this, editorial from Nature, on this very topic, that academic writing is too precious to be automated.

    Writing is thinking,
    On the value of human-generated scientific writing in the age of large-language models.
    https://www.nature.com/articles/s44222-025-00323-4

    Reply ↓

    • Ms:

      As that linked article says, and as I say in my comment above, writing is indeed thinking.

      Or, to be more precise, some writing is thinking. A lot of writing is not thinking, it’s just spewing out words. We’ve discussed this: chatbot output looks a lot like human output when the human is bullshitting and not thinking very hard, kind of like a student who doesn’t now the answer and so just spews out sentence after sentence hoping that something will hit the target.

      Regarding your statement, “academic writing is too precious to be automated,” I agree with Lizzie that nothing is gained by having students dump chatbot-generated texts into their theses and term papers, any more than it would make sense for them to copy and paste textbook chapters or wikipedia entries.

      The point is not that textbooks are “too precious” to be automated. Indeed, it’s easy to automate the production of a textbook: just stick it in the xerox machine. The point is that this sort of copying serves no useful educational or evaluative function.

      Reply ↓

    • Thank you for the link. I quickly read the editorial and was reminded of my most recent manuscript in which I encountered a certain book which I took many quotes from, and which influenced the structure and subsequently the entire way I view that manuscript.

      This all happened when writing, it was not planned up front. It now seems to me that that the final manuscript was partly made possible by coming across the book during the writing process and using some quotes and later on realizing it all fits like pieces of a puzzle from a certain perspective.

      I wonder what you call that process, or how you would view that process, in light of the ideas that writing is thinking and that academic writing might be too precious to be automated.

      Reply ↓

      • Anon:

        That all sounds just fine. I don’t think it’s so fine to quote from other sources without acknowledging the quote. Beyond any problems with rule violations, two problems with plagiarism are: (1) plagiarism can (and, in practice, often does) introduce errors, and (2) by obscuring the source, plagiarism makes things difficult for readers. Basbøll and I discussed this in our article, To throw away data: Plagiarism as a statistical crime.

        Reply ↓

        • The article you link to notes at the end of it that parts of the essay are adapted from Gelman’s blog. This may tie in nicely with what you wrote somewhere else in this comment section about how your blogging may help to explore and organize your thoughts.

          I gave an example in my comment about about encountering a certain book, and using quotes from it, that helped with the structure and such things. Your reply seems to me to focus on the quotes part, but I am still wondering whether there is a name for the process I am trying to describe. Is there is a name for using sources and information and examples you come across during the writing process of a manuscript to make your point, or to make something clear, that you did not have in mind when starting the writing?

          This thinking-writing process I wonder about, and trying to describe, is perhaps similar to me using your comment by referring to the paper you link to and what’s written at the end of that paper, and connecting this to what you wrote in a comment earlier here about how writing helps with exploring and organizing thoughts, and using this all to bring up my question again concerning the thinking-writing process I tried to describe earlier:

          Is there a name for the thinking-writing process where during the writing new information or sources or examples or quotes are used to make a point and/or that fits like a puzzle concerning the thing you are writing about, but wasn’t aware of when starting the writing?

          At this point, to me, it seems more than just finding a (further) example of something, and it seems more than merely combining or summarizing sources or text. It seems more like some sort of interaction between what you started to write, and this new source. And it seems more like some sort of combination of creativity, focusing on certain parts, and connecting things with each other.

          (If there is some sort of term or name for this process, a next question for me would be how this human thought and writing process compares to artificial intelligence writing. More specifically whether current forms or artificial writing can do such, or similar, things)

        • Anon:

          I don’t know, but I have written about the way in which different formats unlock different sorts of writing for me. See here, for example.

        • Ah, thank you for the link to the blog post from 2011 about your writing strategy! Very interesting and useful for me in light of the discussion and what I am trying to describe.

          I think I see some connections with the writing of my most recent manuscript. For instance, you write in the 2011 post:

          “In grad school I moved toward the Latex approach of starting with the template and an outline (starting with the Introduction and ending with Discussion and References), then putting in paragraphs here and there until the paper was done.”

          “When I blog I tend to start at the beginning and just keep writing until I’m done”

          “When I’m blogging I commonly start at one place but then, once I’m halfway through, I realize I want to go somewhere else.”

          “But recently I came up with a new strategy–the best of both worlds, perhaps. I write the outline but then set it aside and write the article from scratch, from the beginning, not worrying about the outline. ”

          I think I recognize similar things concerning my recent manuscript. I had the title and introduction written about two years ago, probably wondered and pondered about the general idea (un-)consciously for a while after that, and started and stopped writing a few times. At one point in this process I made a more general outline with headings and topics and sources to possibly consider further and talk about within these sections.

          And then at one point in time, during the final writing-phase earlier this year, I just began writing in a more go-with-the-flow manner, and wrote the majority of the manuscript in a week or so. It was also then when I encountered the book, and was able to use it in several ways, etc. etc. Perhaps this last part resembles what you write concerning blog posts: “When I’m blogging I commonly start at one place but then, once I’m halfway through, I realize I want to go somewhere else.”. For me it was perhaps more not realizing I want to go somewhere else, but it (also?) kind of sort of happened automatically, as a result of coming across certain sources and information.

          Very interesting to read your post about the writing strategy, thank you again for sharing. I think there must be some papers or books written about writing strategies for scientific papers, but now I wonder if the sort of remembering- and retrospective view you use in your blog post from 2011 might be an interesting and useful addition to the possibly more technical papers or books about writing papers.

27. You say:
    “I have limited hours in my days, weeks and life and thus limited hours to dedicate to graduate student training”

    Which made me wonder what future you are training them for. If they are to work in an academic setting then this may be a reasonable approach (I really don’t have the experience to say) but if the students then move into a world where chatGPT use is encouraged, or even required, then they might enter that world under-prepared.

    Reply ↓

    • Tom:

      I fear that in the near future, lots and lots of people, both inside and outside academia, will have to deal with AI slop of all sorts, including reports that are full of undigested chatbot output. Dealing with bullshit will remain an important skill, and it’s worth teaching. I don’t think the way to teach it is to encourage students to produce bullshit for others to have to wade through.

      Reply ↓

28. I’m a little late to the party but have a few comments:

    For PhD/MSc theses much of the thesis will describe design, implementation, description, analysis and interpretation of “experimental” work and the scope for use of AI here is presumably mostly for grammatical/spelling correction. That doesn’t seem too problematic to me.

    The literature review part of the thesis is susceptible to misuse of AI tools. I think Lizzie’s prescriptions are fine for her and this and related prescriptions can form the basis of what could be Uni- and/or field-specific rules for use of AI. For example there could be a requirement that a student signs a statement that use of AI rules have been followed, that all references have been checked to ensure that these are real and relevant for the citing sentence(s) in the thesis.

    The supervior has an important role to play. Like Lizzie, they can use their own rules/advice about acceptable use of AI (if institution-wide rules aren’t in place or the supervisor feels that these aren’t quite right in context). IMO a supervisor should read and make some suggested edits on written work as well as discussing these with the student during the writing process. I like this part of the supervisory role.

    It’s possible to create marked assignments involving significant writing (e.g. a final year literature review) in which abuse of AI can be minimised. e.g for final year literature reviews (I work in a biomed field) I now ensure that students include some use of bioinformatics software as part of the review. So if they’re writing about rare diseases (say) then I expect them to do some analyses of relevant proteins (e.g. finding and mapping of pathogenic mutants using databases and protein analyses tools and so on). This allows them to demonstrate sklls in methods used in their course and that they might use in the future.

    For me writing is really important (it’s part of how one learns and finds out what one thinks about a subject), and so I make lots of effort to help students with writing, especially those from non-native English backgrounds including some editing of parts of the work to highlight how to deal with problematic grammer, sentence construction, theme development etc. Even if essays and theses are summative assessments there is a strong formativie element to these too and it would be remiss of a supervisor not to address this.

    Reply ↓

    • Chris:

      What you say here makes sense to me.

      Reply ↓

29. In the educational context, oral examination and oral defense of theses have never been more essential. It’s likely that the (formal) importance of the written document will decrease substantially in such evaluative contexts. Anecdotally one already sees this happening in Spain with undergraduate theses.

    It’s hopeless to rely on supervisors to teach good practice when a substantial number of supervisors regularly abuse AI technology themselves in their own work.

    Reply ↓

    • I hope students will continue to be examined in writing, albeit off-line and under proctored conditions. It is possible that something similar also has to happen to the evaluation of scholarship, so that scholars (i.e., supervisors) will have to regularly demonstrate that they can write what they think in their own words. They might have to sit for an examination in hiring, tenure, and promotion situations. Perhaps grant applications will need to be written under invigilated conditions. But that is going to be a touch sell, I think.

      Basically, I hope AI doesn’t undermine our respect for the ability to write well.

      Reply ↓

      • I agree with the sentiment, but this seems too focused on just one aspect of evaluation: to remove the aspect of cheating. I think the ability to collaborate, ask good questions, seek additional information and feedback are also important for research. Your examination protocol could lead to ignoring these things while pursuing the goal of ensuring that researchers understand their own scholarship. These are all important goals and I think we need examinations that include them all with appropriate balancing.

        AI has highlighted concerns about authenticity and researcher understanding. But it also highlights the importance of using many different tools available. The misuse of a particular tool should not prevent its use altogether.

        Reply ↓

        • Yes, my view is that both scholars and students should be free to use AI as they please. They should just *also* be able to put words together to form sentences that express their own ideas. To the extent that AI can generate prose that *could* represent ideas that a scholar *could* reasonably have (if in fact that scholar has knowledge of their subject) we need some other means than merely submitting a paper for publication to ensure that our academics know what they are talking about.

        • Again to continue the analogy with textbooks: Scholars and students should be free to use textbooks as they please. They should also be able to put words together to form sentences that express their own ideas. To the extent that an existing textbook includes prose that could represent ideas that a scholar could reasonably have (if in fact that scholar has knowledge of their subject) we need some other means than merely allowing scholars and students to cut and paste from textbooks to ensure that our academics know what they are talking about.

        • Exactly right!

30. How long until one of these bots can create a persona and get a degree on its own?

    To the extent “talking smart” with lots of jargon can intimidate people into not arguing, its already here.

    Reply ↓

    • Anon:

      If the bot plays its card right, it could definitely slide into a job as Dean of Engineering at the University of Nevada. And if there’s a concern that the bot is “hallucinating” and making up numbers, that would qualify it for a tenured position at Cornell or Duke, or a job in whatever part of Columbia’s administration is tasked with providing numbers for its U.S. News statistics. And if all that is required is the ability to ramble incoherently and make stuff up, then it could compete convincingly for the job of president of the United States. So, yeah, a simple college degree would be the least of it.

      Reply ↓

      • In a sense, AI makes it easier to spot such incoherence or deception. With humans, they can sound convincing sometimes – LLMs still have that recognizable computer identity. Example: I was listening to a Republican press conference today and heard one speaker attacking the Democrats for voting against the great bill. She said that Democrats should be held accountable for their vote to raise the tax rate for someone earning $48,000 from 12% to 25%. She claimed this was a “fact.” It sounded pretty dramatic and she was well spoken (although self-righteous). So, I looked up the data to check this “fact.” It turns out to have two problems. First, she is talking about marginal tax rates and I suspect the majority of the public doesn’t understand the difference between marginal and average tax rates. Second, the current marginal bracket of 12% would have gone to 15% without passing the bill. 25% is the next higher bracket. But she sounded quite convincing. LLMs may have convincing content, but thus far still have an artificial voice (sound and/or writing style).

        Reply ↓

        • Dale:

          In this case it sounds like the problem was not so much with the details as with selection–as if the impact of a complicated tax bill could be summarized by a very specific number regarding the marginal tax rate of someone who happens to be very close to a threshold. I’m reminded of the research studies we often criticize, where the result of some complicated study is summarized by some very particular comparison on some subset of the data that happens to exceed a statistical significance threshold.

          I’m thinking that part of the problem here–beyond political speakers who are willing and able to mislead and, where they deem necessary, lie–is the mistaken intuition that everything will go in the same direction. So if the bill is a good idea, it will only have good effects, and if it’s a bad idea it will only have bad effects. We see this attitude a lot in junk science (for example, in a book like Nudge where every study they mention just happens to fit right in with their story) and implicitly it’s here in this sort of political argument, where it’s just accepted that you could summarize the effects of a big tax bill by this one statistic.

          People are aware of this problem–there’s an expression, “cherry-picking,” which applies here–; still, I think there’s a problem that these cherry-picked statistics can derail discussions of the big picture.

31. These discussions always trigger a Johnny Cash track in my head:

    YouTube: The Legend of John Henry’s Hammer

    To borrow a quote from Jeanette Bickell from the Wikipedia page, John Henry (folklore), “John Henry is a symbol of physical strength and endurance, of exploited labor, of the dignity of a human being against the degradations of the machine age, and of racial pride and solidarity.”

    Reply ↓

    • Another thing this stuff reminds me of is Luddism. which most people are completely wrong about.

      https://pluralistic.net/2023/03/12/gig-work-is-the-opposite-of-steampunk/

      “As Brian Merchant documents in Blood in the Machine, his stunning, forthcoming history of the Luddites, the factory owners of the industrial revolution wanted machines so simple that children could work them, because that would let them pick over England’s orphanages, tricking young kids to come work in their factories for ten and twelve hour days.”

      There’s a lot of similarity in the social conditions surrounding all of this.

      Reply ↓

32. I am not a fan of students using LLMs to compose their writing, in part because I see that they are cheating both themselves and our overly credentialist society by doing so; however, if a graduate program certification can be cheated by someone employing an LLM to write their publications, it already can be and, thus, has been cheated by someone with enough money to hire someone to do their work for them.

    Reply ↓

33. Thank you for doing this.

    Reply ↓

34. I really appreciate how thoughtfully you’ve laid out your reasoning here, Lizzie. Your perspective highlights an important tension in academia right now: balancing authentic student learning with the convenience of generative AI. Many supervisors share your concern that students may miss out on the deep writing and critical thinking process if they rely too heavily on tools like ChatGPT. Writing is not just about polished grammar it’s about learning how to build arguments, develop clarity of thought, and communicate effectively.

    Interestingly, this discussion reminds me of how media and streaming technologies have disrupted traditional entertainment. Just as academia is figuring out how to set guidelines for AI use, viewers are also navigating a shift from conventional TV to IPTV services. Platforms like Magis make that transition smoother by offering access to live TV channels, movies, and sports streaming in one place but users still need to learn how to use these tools responsibly to get the most value.

    In both cases whether it’s students writing theses or people choosing how they watch television the key seems to be intentional, transparent, and mindful use of technology. AI and IPTV aren’t inherently negative; they just require clear boundaries, whether it’s institutional policies in education or choosing trusted platforms in entertainment.

    Do you think universities should treat AI in student writing more like how publishing treats plagiarism policies with standardized, enforceable guidelines rather than leaving it up to individual supervisors?

    Reply ↓

35. Thank you for doing this.

    Reply ↓