Regardless of how you feel about generative AI—love it or hate it—the worst thing you can do with it is use it to write for you.
Imagine, for the sake of argument, that there was an LLM entirely free from any ethical, ecological, or economic baggage: it had been trained on public domain text using fusion energy and could run inference on Raspberry Pi-tier hardware. Everyone in the world had access to it, it never hallucinated, and was fully aligned with doing good for humanity. This would be very exciting and useful—I’ve found LLMs to be amazing for translation and accessibility and having a cheap, ubiquitous, and reliable tool to aid those goals would be phenomenal.
It would still be a bad idea to use it to write for you. Why?
Writing is thinking
I am not the first person to suggest that writing is a critical mode of thinking. Rose Horowitch writes in The Atlantic:
Writing is hard. Orwell likened the experience to a “long bout of some painful illness.” AI promises a simple remedy. The trouble is that writing is not merely the act of transcribing fully formed thoughts—if it were, it wouldn’t be hard. Writing is the way people figure out what they think, and how to convey those thoughts to someone who doesn’t already share them. Cal Newport, a computer-science professor at Georgetown University, argues that the process of writing forces people to think in an orderly, linear fashion. It exposes flabby thoughts and shoddy reasoning. And the time and focus it takes to form thoughts into words, sentences, and paragraphs allow the author to make new connections and discover new insights.
This feels true to me. My job is to write. …The writing process is how I refine and formalize inchoate ideas and gain new understanding. By evaluating my arguments and discarding those that aren’t convincing, I find the ones that are. Writing is hard because the writer is learning. If AI eliminates the challenge, it also eliminates the learning.
— Horowitch, The End of Reading is Here
I agree entirely with that passage. When I have a hard problem at work, the best thing for me to do is to describe the problem in writing and then to “narrate” my thought process on the page a little. This helps me to know what I think about something.
Writing is hard. I’m struggling to write this blog post right now. I keep revising and revising in my head and in my editor—and that takes work. But at the end of this I’m going to be better able to articulate exactly why I think using AI to write something for you is stupid, and I’m going to have a much clearer understanding of it myself. I could have, of course, dropped a prompt into $LLM_OF_THE_WEEK and gotten a polished essay. I could have read it carefully, nodded my head in agreement—but my mind wouldn’t be the same.
AI is like a forklift: it can be steered by a single individual and accomplish tasks that would have hitherto taken a team. That’s great. But we don’t take our forklifts over to the benchpress at the gym and start busting out reps. The gym is for us to improve, and that takes personal struggle. Writing changes our brains and that is half the point. (The other half being able to communicate your idea to someone else, of course. This is optional in many circumstances!)
Writing fluff
I get that there’s a certain amount of writing that people do in the course of their work that’s just busywork: maybe your boss wants daily status updates on a project or something a rather. I still think it’s better to write by hand. Maybe it’ll be quick and sloppy—that’s fine, an LLM is better employed in cleaning up prose, but not composing it in the first place. If a human is going to read something, I don’t think there’s any excuse to have a bot do the writing.
Last year I had an experience with “work slop”, though that term hadn’t been invented yet, I think. It was awful. It was degrading and disrespectful. Foisting slop onto someone else is terrible.
Programming can be writing sometimes
With that in mind, I feel like programming can function like writing prose sometimes. If I’m sussing out a new problem domain, programming by hand helps me get intimately acquainted with the terrain. A lot of code written in the corporate world—at least in my experience—has been pretty uncreative and frequently tear-inducing boring. Most code exists to expose business patterns and protocols to automation, and there is a lot of overlap between companies in what kinds of things need to be automated. LLMs excel and writing this kind of stuff.
But there’s still that piece here and there where the domain isn’t well-known to the people maintaining the software. These are the places where working something out by hand can be more beneficial than running blindly ahead with an LLM because of the understanding that you build up in your own mind. You might even encounter new information that changes how you see and approach the problem entirely. I’ve seen this happen in my work, and I sure I’m not the only one to experience this.
Minds and machines
I’ve been impressed with some of the coding that an LLM-backed agent can do. I will still write code by hand when I need deep contact with the problem domain because it helps me think.
I think writing with an LLM is completely stupid. Writing is thinking. If something isn’t worth thinking about enough to write manually, it probably isn’t worth writing at all.