Scientific Writing Is a Miserable Grind, So Let's Outsource It to AI

Very little of the science publishing process is valuable ⚗️

The Alchemist, by David Teniers the Younger (circa 1651-1656)

Scientific publishing is broken in all sorts of fun¹ ways, but the most tragic has to be how miserable the entire process is for everyone involved.

Everyone agrees that writing papers is necessary to be a researcher, but nobody really enjoys writing them. The only accepted styles are deliberately dry, out of a pretense of objectivity. When a junior researcher tries to add a bit of fun in their work, they are told that they should strive to make their papers boring. Countless grad students have been known to put off writing their theses, in theory their life’s work in the area they’re most passionate about — and though this is only speculation on my part, I’d venture a guess that it’s because they secretly (or not so secretly) hate the writing part.

Once the writing is done, the paper is submitted to a journal or conference. Leaving aside the fact that this is typically done through the least user-friendly web portals known to the modern internet, the submitting process itself is soul-crushing in several ways. One of them is the insane delays: the time between submitting and getting an answer (which may of course be a negative answer, prompting the researcher to start over with another journal) can take months, during which the research can become outdated. Sometimes you even run into the fun² catch-22 of having a reviewer complain you don’t cite references from the current year when the paper has been “in review” since the end of the last calendar year.

Stu Maitland @StuMaitland

Just had a paper rejected; one of the reasons was "not using references from 2022". Dear reader, the paper has been in review for over 6 months

Speaking of peer review, there are few things out there that are as broken and hated and yet simultaneously defended as if it were one of the few bulwarks of civilized society. Peer review relies on unpaid labor from researchers all over the world, who have been led to believe they must read anonymous paper submissions as part of their “academic service.” Obviously they don’t enjoy it. It’s more of an annoying and recurring obligation. Nor are the authors who receive the review any happier. It has become a cliché to complain about “reviewer 2” and their senseless, harsh comments. Albert Einstein famously submitted a paper to a peer reviewed journal only once — this was a different era — and he hated it so much he never did it again. How many among today’s scientists would follow Einstein’s steps if we hadn’t enshrined peer review as an almost sacred thing?

All of the above could be necessary evils if they were producing something of value. Other forms of publishing, from journalism to short stories to nonfiction books, also involve strict stylistic constraints and dehumanizing bureaucracies. But they differ from scientific publishing in that their output is something that people are generally happy to read.

Nobody is happy to read a scientific paper. Sure, you might be interested in the contents. Sure, if you’re a researcher, you do want to be aware of new developments in your field, and reading papers is a way of doing that. Sure, sometimes you find a rare example of a well-written paper that is a joy to read. But let’s be real: the median experience of reading a research paper is a miserable one. Most of the time we default to reading just the title. Maybe the abstract. Even most abstracts are painful to read, their only redeeming feature being that they’re short.

And that’s if you can even access the paper you want, which most people outside universities can’t unless they pay ridiculous sums of money.

The only true winners of the scientific publishing game are prestigious journal editors. Everyone else — authors, reviewers, readers — loses. And broader society, which should be the main beneficiary of scientific advances, seems to be losing more and more as science slows down or even produces dangerous research due to the publishing incentives. Scientific publishing is a system that takes human misery as input and produces human misery as output. It might be time to change that.

Last year, I spent a lot of time thinking about AI-generated art and its consequences on artists and our relationship to beauty in general. Discourse on that topic seems to have calmed down recently. I’m not sure why, because progress in AI art sure hasn’t stopped (see for instance the release of Midjourney 5). Maybe we’ve all somehow accepted that this is the world we live in now. Maybe there have been too many other developments to pay attention too. Maybe we’re just waiting for the outcome of a few lawsuits.

If you haven’t followed along, Midjourney 5 is now pretty good at realistic hands! Still missing a finger on this alchemist’s right hand, though. We’ll pretend he lost one in an accident involving a blast furnace or something. (Thanks to @TheGrandBlooms for generating it!)

One big reason why AI art is controversial is that art, by definition, is something we humans do because we like it. Both the input and the output are enjoyable. If AI “steals” the joy of learning to draw and making cool pictures for others, and ends up weakening one of the core purposes of human existence, then almost everybody would agree that that’s a problem. The true controversy is upstream — on whether AI actually does that.

We haven’t seen an equivalent level of controversy over AI-generated writing. I’m guessing it will come soon enough, now that GPT-4 probably matches the skills of a lot of human writers. Is it bad if we can just generate a great short story from a short prompt? It certainly is bad when spammy, low quality AI-made submissions clutter the submission process of a short story magazine, but what about that point in the near future when the AI stories actually become good?

That’s a debate for another time. For now, it’s important to note this one fact: unlike visual art, there are two kinds of writing. There’s writing that we write for pleasure, or for other meaningful reasons like exploring an idea. This category includes most fiction, poetry, essays, and blog posts. And then there’s writing we write to fulfil utilitarian goals, such as making money or advancing one’s career. This includes most clickbait, marketing, news, corporate reports, and, of course, scientific papers.

It seems to me that we should think carefully about automating the first category. I want to keep writing blog posts, and I wouldn’t be happy if it became pointless to do so because of GPT-5 or whatever. But the second category? The one that mostly involves misery on one, if not both ends of the pipeline? Automate away!

Scientists have already begun using ChatGPT to write their papers, to the point that some big journals have been clarifying that ChatGPT cannot be listed as an author or even banning the use of AI in submissions altogether. The first point makes sense: being an author implies more than just having contributed words. After all, the word author shares an etymological root with authority: an author should be ready to defend their work, a responsibility that ChatGPT can’t bear.

But the second point seems exaggerated to me. Of course, the suspicion exists for understandable reasons: AI-generated writing is full of inaccuracies, misleading statements, and even hallucinated academic references…

ChatGPT is perfectly able to confidently tell you the abstract of an article that doesn’t exist. (Source)

… but none of that matters if you just use AI as a tool whose reliability is understood to be imperfect. It would be stupid to ask a large language model to generate a paper of your research and then submit that as is, without so much as a glance, to a journal. It’s not so stupid to ask it to rewrite your awkward prose into something more easily digestible, verify the content, edit it some, and then submit that.

In fact, that’s probably something that researchers should all already do, if they’re not submitting to journals that have banned LLMs. I don’t use LLMs in my blog posts because I enjoy the craft, and also because I both trust my own abilities and want to develop them further. But if a scientist who doesn’t enjoy writing, isn’t that good at it, and doesn’t care that much about getting better, she should use all the tools that allow her to easily produce something that others will enjoy reading. That’s less misery on both ends.

Sure, maybe the scientist would benefit in other ways from writing her paper herself with no help from AI. Maybe it would be better for understanding her own work and ideas. Maybe she would be more likely to spot mistakes in her research. But honestly, that should be up to her to decide. She can still (and maybe should still) write an informal blog post about her research, which will be more fun for both herself and others who are interested in her work. The LLMs can take care of the tedious, formal part of publishing.

We just need publishers to accept it. That might be difficult in some cases, since big science journals rely quite heavily on conservative norms to keep their prestige advantage, but maybe AI is what will finally force them to adapt. There are at least signs that some of them see the adoption of LLMs as inevitable.

Eventually, large language models will get good enough that they can create scientific papers of perfectly good quality with little need for human supervision. At that point, we may not be very far from just automating all of science — what Holden Karnofsky calls PASTA, or “Process for Automating Scientific and Technological Advancement.” PASTA, if it happens, is likely to be a totally transformative technology. It will turn our world into something that might be very difficult to recognize, so it’s hard to predict anything on that future timeline.

But for the (maybe short) period before then, we will be free from the soul-crushing industry of science publishing, and it’ll be amazing.

The Alchemist, by David Teniers the Younger (circa 1652). Yes, same title, same artist, and same approximate year as the painting at the beginning of the post. In fact there’s a whole Wikimedia category called “Alchemists by David Teniers the Younger”. That must be why Midjourney is able to generate “an alchemist reading a book in the style of a 17th century Dutch painting” so easily.

1. Not actually fun.

2. Also not actually fun.