How I use AI for academic writing (while avoiding AI slop)
Because I work in Digital Humanities and part of my job is thinking about the ethics behind AI, I get asked a lot how I use AI for academic writing. Today I finally want to publish a brief post about how I do it. All in all, there are three big things for me that have lasted the test of time: two of which are custom GPTs, and one is an app that I use for “reading”, called Listening.com.
My uncommon approach to using AI in research
I have played around with AI in academic writing a lot, and when I have described my process to fellow academics it seemed to resonate with them a lot. Yet I suspect it’s a bit different from how I think a lot of people use AI (in academic contexts and elsewhere).
For instance, I don’t use it for summarizing PDFs or things like that. Really nothing that lets the AI do the actual work or actual thinking. I want to do that work myself, because it’s part of this engagement with the scholarship before me and I need to know what’s actually in a text first hand; otherwise I would spend excessive time anxiously looking for hallucinations or mistakes (and would be right to do so!). And yes, often AI summaries are quite okay and if you need a quick summary, fine, but I feel like the AI often loses the most interesting things. I’m ultimately interested in random stuff, like scholars are. So yes, it’s nice to have the key argument of an article, but that’s what an abstract is for. If you want to know the key point of the article, reading the abstract usually does wonders for you, and maybe the introduction and conclusion, like academics have been doing for a long time.
Please don’t use AI as a search engine. The quality is awful, yet that’s often hard or even almost impossible to tell when you’re not an expert on the topic. The more the information you’re looking for is part of a knowledge mainstream well represented on the internet, the more likely the results will be somewhat ok. But when does that ever apply to knowledge relevant to scholarship? We don’t want the same old problematic, unscientific shit people say on the internet (even though, to be fair, AI training contains lots of pirated work from our scholarly peers – as long as it’s just the exploitative publishing system they’re ripping off, I wouldn’t mind).
Also, did you know we’re all contributing to environmental degradation and rewarding awful inhumane, exploitative, super unethical working working conditions by using AI tools? Just something to think about before you continue.
In any case, I use AI for my research exclusively in ways that are less prone to hallucinations and that leave all the scholarship to me. I don’t want AI to do my research. Jobs in academia are scarce enough. AI slop is everywhere. We don’t need any more of that. I want to access unique ideas of actual humans, not AI suggested regurgitations of (potentially problematic) mainstream thoughts found on the internet. I want diversity. I want marginalized voices. I want subjugated knowledge from the periphery, not just the center of dominant worldviews. I don’t want AI to write my papers for me and do all the thinking, then all I get to do is laboriously check for errors. That’s not how it’s meant to be. That doesn’t improve my life. So here’s what does.
Listening.com for academic “reading”
Over the past years, I have come to realize that if I want to get a considerable amount of reading done, it has to be in audio book form. I spend so much time reading and looking at screens, it simply doesn’t work for me to look at physical books in a regular-enough fashion to stay up-to-date with my reading.
I have tried Speechify in the past, but I like it a lot less than Listening.com, even though I don’t appreciate that Listening recently removed German as a supported language. Most of the literature I use is in English, but every once in a while it would still be practical to read one of the older articles in the German tradition as well, and I don’t understand why they removed it, because it clearly worked in the past. Anyway, I’m getting ahead of myself.
I have a very particular and peculiar way of looking at things (in general and other people’s writing in particular) and that’s what makes everybody’s research unique. So I prefer putting texts I want to work through into Listening.com. I listen to most of my articles these days, and it has drastically increased the amount of reading I can get done without actually feeling like more work.
Some people argue they’re much faster sight-reading than listening to something (which is a good point) but you can speed it up if you want (with some practice, you can gradually crank up the speed quite a bit). And Listening.com allows you to listen while you’re doing something else, during which you couldn’t sight-read. That’s the key advantage for me. It helps when your eyes are tired. My problem is not that I’m a slow reader (I’m actually a very fast reader, I’d like to think). But looking at screens is strenuous (for head and eyes) and when I work with print-outs, I often don’t sit very straight (which is bad for my back pain with which I have lots of issues). I love not being always glued to a screen while still being able to “get some reading done”. I use it while biking to work or while going on a walk (on my walking pad or during commute times). This helps me make the most out of otherwise “dead time” and I find that in these calmer moments, especially where my brain is stimulated by some movement, I’m far more open to actually listening to text.
Of course, I still skim-read when I need to add an extra reference during a writing session. But the good thing about Listening.com is that I also get times where I can “read” slowly and think about things more deeply. Sometimes it’s nice to get an extra boost of skim-reading info during a writing sprint or to look something up. But if you ever only do that, you’ll never generate new ideas. (I don’t want to speak for you, maybe it’s different for you but I know that I need to spend some time with the literature and do so on a regular basis to generate new ideas). My AI reading tools help me do that. I can do academic writing, AI-assisted, without being glued to a screen, which is worth a lot in my book.
You can also use those apps to re-listen to your own writing after finishing a draft or during editing phases, if that’s something you like to do. I do that every once in a while.
In the beginning, I used to use the app’s “notes” feature to summarize articles. I’ve actually stopped doing that. That used to be a lot of work. So now, if I want to keep a lot of notes or excerpts from the text I just listened to, I’ll just speak the details I want into my AI voice to text workflow to make an excerpt (more on that process later). When I do that, the resulting material is already much more digested and thus more directly usable in my own articles. It’s less of a general-purpose summary, because ultimately what you want is your own reading notes, not some generic bullshit. Even though I do admit that I can imagine cases where you could use “ask my PDF” features to playfully think about the article. If that works for you, who am I to judge? Do whatever floats your boat.
Using AI for my actual writing
To help with my academic writing, I have this custom GPT called “fix academic writing.” I will speak my input text into it (using my phone’s microphone and the app’s speech to text feature), which is often not complete sentences, and AI will put out something that is basically the same text, but nicer, and that’s already a usable base draft with complete sentences. This really helps you get rid of the fear of the blank page, because I talk a lot, so that’s easier for me than writing by typing (which may not be the case for you but I suspect it is for many people). The resulting first draft is just so much cleaner than what I would produce by hand, because I’m not a native speaker and I sometimes use structures that aren’t quite correct. When I do this AI-supported, I already start with a draft that’s much more workable than what I would otherwise create.
I think this is really a game changer for people who are not native English speakers. If I want my output text to be in German, I sometimes still do the initial version in English, because I work solely in English nowadays and the AIs are simply better in English. Then I’ll create the German output text by translating it. I’m a bit wary of using German as the prompt language, to be honest.
As for these transcriptions from my voice input, I feel like the AIs are getting more and more loose with that, so definitely check what the transcript actually says and, if necessary, fix before sending it off. It does tend to make mistakes when it doesn’t understand something correctly. For example, with my early modern alchemists, it sometimes uses background knowledge to “improve” things. Nowadays the transcribing part has a lot of AI in it, and sometimes it really does include things that I know I didn’t say (adding some references to alchemy being “esoteric” is a classic). So you have to be careful—not because it’s wildly wrong, but because it may use formulations that you didn’t use. I use this voice-to-text transcript (used as input to the text fixing LLM) as a kind of ground truth, to make sure that the summarized text is really what I said later on (as I write I often only give this a cursory look to check for major red flags). When using it to summarize or excerpt articles I have listened to (as explained above) that’s often the only excerpt I’ll have from the article (beyond the article PDF itself that I can, of course, always consult again later), so I make sure this is mostly ok.
Maybe once a month (or earlier if I directly need my draft texts for currently ongoing academic writing) I transfer these draft texts (including a link to the AI conversation that created it) in a text document, often including my voice input transcripts. These documents then get printed out, as I’ll likely need a quick overview of the material again during later writing stages and that simply works better from paper, if you ask me. Often, these very raw drafts are what I call “pre-draft material” that I don’t think is fit to be integrated into articles directly. If I were to do that, that usually ends up needing massive amounts of editing and often, it’s cursed and will never truly work. I fix that by doing a new voice-to-text-based draft (as described here) by speaking what I later find to be the key aspects into my phone again. This second version is usually much more “digested” already and more directly usable in my text.
Nowadays, I’ve even started keeping only the link to the chat and the processed text after a brief sight check right after it was produced, to make sure it’s not outrageous. Later, in my notes, I’ll correct the details that aren’t right, because there are usually details that are misrepresented in some way. This is also a good time to remove obvious AI-writing artifacts (like words that are unusually common in AI-generated writing). All in all, this is a really fast way to get good basis material, but it usually ends up being too much basis material that is somewhat redundant. To deal with that issue, I have another GPT that takes all this redundant basis material and it creates an article section from selected material (I usually give it a bit of a hint about where I want the section to go).
If I really need all the details from an article (I had listened to and excerpted) again, I’ll just read it again. The ones I consult and cite a lot, I’ll need to go through again before publication anyway, to check whether I summarized them correctly. So it’s not really something that saves me a ton of work per se. I think it just makes the work more comfortable for me and I find I produce better work like this. I wouldn’t say that using AI makes me that much faster. I can be faster in producing a midway okay text to submit if I’m really late for a deadilne, but for the actual work, I think it mainly helps me do it better, because this is a quite complicated workflow.
One more tool that sometimes helps make writing faster is what I call the “text to academic writing section” custom GPT (already hinted at earlier). Basically I will input draft material created in the step before (ordered by topic or section this will be discussed in, along with some instructions on the line of argument in the section) into another custom GPT with its specific instructions and the result will be a first draft of a section based on my input material. This can usually be edited into a truly usable section for an article in a reasonable amount of time.
You have to understand that I talk a lot and tend to be a bit redundant when I’m still in the process of thinking about something. When I speak into my phone, I repeat myself; when I keep working on articles and listening to articles, things get redundant. Not only because the research literature itself is a bit redundant (once you’ve reached the point where you feel like everything repeats itself, that’s where you’re ready to write the literature review) but also because I’ve already repeated similar things multiple times. Then I don’t remember that I’ve already inputted certain text, and I do it again. LLMs also tend to create more and more text, so you find yourself in the situation where you need to reduce it again to only keep the most relevant stuff for your article .
If it’s a lot of input material, you really need to make sure the output truly contains everything you wanted said. But I actually like that it sometimes leaves things out because I have a really hard time letting go of material. If you’re the opposite, maybe that’s different for you, but I produce too much text, and once it’s produced, I have a hard time letting go of it. This custom GPT has really helped me get rid of unnecessary text and overly long excerpts. You don’t need a 1:1 copy of all the articles you’ve ever read in your own words in your literature review; the purpose of most sections is to be brief and only keep what’s most important to your argument. I really struggle with that without the AI.
So this is how I use AI in my research. This workflow is so helpful for me specifically because it works for me. You can tell from my workflow that it’s really tailored to how I function and what my specific challenges with academic writing are. An optimal solution might look different for you.
Yet, I have found that especially how I do my excerpts and how I produce the first version of a text by just speaking it into my phone resonates a lot with people.
Importantly, I currently love doing my writing like that. It removes many sources of procrastination (like the fear of the blank page) and counters my specific writing challenges (such as embedding my German thinking into English text in ways that become almost unfixable without rewriting it completely and the fact that I have a hard time letting go of details).
I nowadays create drafts for my blog posts in a similar workflow (but one where the initial voice-to-text scenario retains more of my uncensored way of saying things, because I know you love my blogposts for being blunt).
Of course, the intermediary output of such a workflow will always be a little bit AI-sanitized, but you can mitigate that by how the prompt is set. And, of course, you need to go over all material yourself anyway to check if everything is correct and said exactly like you want it.
All in all, I enjoy this process because it brings out the best in available AI solutions (as deeply flawed as they are and simply finding a good workflow still isn’t a great excuse to use them anyway, unfortunately). What I like about it is that it is relatively safe. I’m never relying on the AI’s “knowledge”. It sometimes hallucinates but everything is based on material that I spoke into my phone, so it’s relatively easy to spot when it diverged and the divergences in this relatively controlled environment usually aren’t that crazy.
I’m sure it still sometimes influences my thinking, especially when I use it to re-write sections (that custom GPT is a bit more creative than the other one that I only use to fix the grammar) but I hope it’s not too much and I will go over all text produced that way many times before it gets published, even if it’s just a blogpost. The AI is a relatively easy way for me to fix bad non-idiomatic sentences, which is a big plus for me as a non-native English speaker. Am I still critical of AI? Yes, very. Do I want to follow through and let go of this workflow for ethical reasons? No, definitely not. Maybe I could finally take the leap and move to better AI providers. But change the workflow completely? No. This is how I work now and it works really well.
I hope this blogpost was useful to you.
How do you integrate AI into your workflow?
What are your experiences and thoughts?
Best,
Sarah
OpenEdition suggests that you cite this post as follows:
Sarah Lang (February 15, 2026). How I use AI for academic writing (while avoiding AI slop). Epigrammetry. Retrieved March 14, 2026 from https://doi.org/10.58079/15ozl
