Recent months stirred up the landscape of human-generated content, from images to presentations to text. Recently, I came across a Twitter thread about AI-based tools generating fantasy novels. My colleague’s (whose quips are always epic) tweet:
Sure, you do not want to write the umpteenth motivation letter or memo. But as there is more to art than the end result (refer to P2P No. 25 — Kill your darlings), so is there more to writing than mindlessly using such tools?
That is, you need to think about your goal. What you write is content (what you want to say) and style (how you say it).
Content
“Writing is the process by which you realize that you do not understand what you are talking about. Importantly, writing is also the process by which you figure it out.”—Farnam Street
You should know what you want to write about, hopefully in the beginning, but at least in the end. That is, it is your responsibility to generate the content—I am not using content as in “content creator”; I am using it as the crux of the writing, the essence, assuming thoughtful consideration of what you want to say.
It is tempting to simplify your life by using AI to generate your thoughts (let this sink in for a moment),
but if you can avoid it, then do avoid using ChatGPT and the like for such. If you are unsure, repeat in your head: “I am delegating my thoughts to some computer code.”
Of course, there is nuance to this as well. If you collect your thoughts into a list but struggle to weave them together, then AI-based solutions can help. E.g., connecting the dots for a project pitch can be eased by a generative tool. This is bordering between content and style, you might say, since it is given what we want to say; only the how is missing. I agree to a certain extent, though I think the logical connections between thoughts have some content-like qualities.
Style
It would be unorthodox to see news articles, scientific papers turned into bullet points. Everyone knows people who, when they start talking, get into sentences Prust might envy. Then the content gets lost, which is the tendency in a non-negligible amount of academic papers (besides hedging like “might possibly suggest”)—they do not optimize for clarity. They often sound pompous by unnecessarily sacrificing brevity to suggest they are more “sciency”. Adding equations just to look more complicated is one point. If you do not have the fortune to learn science writing skills and will try to pick up how to write papers from the literature, then you will gravitate toward this style. I did it, too (hopefully not so much now).
If you get feedback that your writing is hard-to-follow (reviewer two might even be right to some extent), then AI tools can help. You can ask to edit your abstract, to rewrite it for brevity, clarity, or to remove hedging. You might even try to rewrite it to a style you find convincing. You can get suggestions on expressions, word choice, and proper tone. This is where these tools can excel—engineering good enough prompts is an art, but trial and error can go a long way.
What I like about these tools (besides being fascinated by the technology as a researcher in AI, though not in language-based models) is that they make experimentation and finding your own style much easier. You do not need to rewrite text to see whether a specific style works. And though the generated version's content will not perfectly match what you wrote, it can help in exploration.
If you want to generate text, you might generate some fancy lorem ipsum.
Ideation
I think about these tools as an aid for extending our creativity. I know how hard it is to come up with a concise and catchy title for a paper: do I want to pose it as a question? Or make it summarize the conclusion? Or should I describe the topic as precisely as I can?
By giving context to large language models, you can ask to generate ideas.
Your own editor
If you have ever felt overwhelmed by the intimidating page limits of scientific papers, spending ten minutes frantically searching for a synonym with two letters less, AI-based tools might also help you. As Ethan Mollick demonstrates in his posts, these tools are powerful editors.
Of course, AI is not a human editor, but it is more affordable.
If you get lucky and your not-so-well-edited papers slip through peer review, then it is easy to anchor to that experience and think that it is enough. But it is not. I will be blunt and say again that clarity beats everything in scientific writing, at least it should.