_                   _                      _   
 _| | __ _____ ___ ___|_|___ _____    ___ ___| |_ 
| . ||. |     | . |   | |_ -|     |_ |   | -_|  _|
|___|___|_|_|_|___|_|_|_|___|_|_|_|_||_|_|___| |  
A stream of consciousness.

Producing slop


I find in order to learn anything new I need to attempt to do something with it. For reasons I can’t easily articulate I decided that I should get into emacs (the file editor, not the discontinued education-focused Apple all-in-one) and figured that using emacs to write this blog would be a good project.

Emacs is famed for its modularity, and of course there was an emacs package specifically for managing Hugo blogs.

Visiting the GitHub page for the package I saw there was an Issue attached to the package that was along the lines of “I have forked this package to use ChatGPT to write blog posts and it’s awesome and you should check it out”.

It is fair to say that this left me flummoxed.

I have always enjoyed writing. One of the reasons I put together this blog was to give myself an outlet to do a bit of writing as my current job doesn’t have a lot of creative writing opportunities.

I have never had “writer” or “author” as a job title, but I’ve done a lot of writing in my professional life. I have written articles that have appeared in peer reviewed academic journals and in edited books, had the odd thing published in newspapers, and at this point have written probably hundreds of thousands of words that have appeared in reports published on websites.

I wouldn’t call myself a great writer, but I think I’m competent. But I certainly would not confuse my enthusiasm for writing with any sort of inherent talent.

Writing is like any other skill – some people are naturally incredibly talented but most people require practice to get good. And unlike a lot of other skills, writing doesn’t provide much feedback by itself. You can tell you’re not a good skier because you fall over a lot and crash into trees. You can tell you’re not a good painter when the picture on the canvas at the end of the process is nothing like what you’d intended to paint. But if you’re at least reasonably literate and you write something that is comprehensible you might consider you did an okay job of it.

Reading a lot can help you become a better writer, but reading in itself doesn’t make you a good writer, like listening to music won’t make you a musician. It can teach you some new words, or some new techniques, or give you some ideas you didn’t have before, but only if you already have the basic skill set to build upon.

The only thing that can help you become a better writer is to do lots of writing and then do a lot of re-writing of that writing.

You can get some of the way by re-reading and editing your own work. But you really need to have a bit of space between writing it and reviewing it, or you’ll just read what you expect to be there rather than what is actually there.

Personally, I hate editing my own work. Maybe everyone does. It’s so tedious, and I’m lazy.

Whenever I write something and it’s garbage or full of errors or typos it’s because I’ve been too lazy to proofread and edit it. This is particularly true of my social media posts, which in my head are ephemeral and it’s not worth slowing down for re-reading or editing, but which in reality just suggests to everyone who only knows me via my social media posts that I’m an awful writer and I can’t spell.

The real secret to becoming a better writer is to have someone else read your work. It’s also terrifying and occasionally infuriating. But it works; it’s possibly the only thing that really does.

I could tell anecdotes about bosses who sent me back documents full of tracked changes and how enraging or dispiriting that could be, but everyone who has ever written has a bunch of those stories and they’re all pretty much the same. Spoiler: it sucked at the time but it helped make you a better writer (or helped make you decide to find a different job).1 We laughed; we cried; we grew as people and still to this day harbour a lingering resentment. There are no special and unique snowflakes here.

But this wasn’t supposed to be a “how to be a writer” post. This was me being flummoxed that someone would not only think that using an LLM2 to write a blog post was a good idea, but was so good an idea to be worth bragging about.

I’m not going to deny that there are legitimate use-cases for LLM-generated slop. Spam is a good one. Or spam-adjacent things like marketing copy, where you write it in the knowledge that no thinking person will be required to read it. Sometimes words are required that it would be unreasonable to ask a sentient being to write.

But using an LLM to write something useful, something where you’re trying to communicate something, just seems so misguided.

If you’re so unconvinced in your own writing skills that you think a computer can do a better job, chances are you aren’t well placed to identify when it doesn’t.

When I suggested this on Mastodon I got a bit of push back from people arguing that even people who can’t write (I quite specifically never said can’t write) can identify bad writing. And while this may be true in the abstract, I think it’s self-evident that if you look at the extremely formulaic slop produced by an LLM and think “that’s fine”, you really can’t identify bad writing. Or you can but your standards are so low that you let it slide.

Chances are if you think LLM-produced writing is anywhere near good that’s more the Dunning-Kruger Effect at play than anything else.

But the real tragedy is that people who use an LLM to do their writing will never get any better at it. No one is going to be inspired by LLM-generated text because it’s by definition just average. If you learn any new words from an LLM, chances are your vocabulary is pretty limited, or the LLM has just made the word up. It’s literally a sophisticated statistical model of what word comes next; it’s necessarily uncreative and unoriginal.

Someone suggested to me that an LLM was okay for inconsequential writing, like writing an email to your boss or a lawyer.

I’m not a lawyer, but I do deal with them a lot. Words are everything to lawyers. The law is effectively the process of using words to define reality. Legal communication does involve a lot of what appears to be boilerplate, which kind of suggests you can fudge it, but can I respectfully suggest you never ever put anything in writing to a lawyer unless you’re absolutely and completely sure it means what you intended it to mean. I really don’t recommend using an LLM for this.

And in my 30-something years in the workforce I have never agonised over writing more than when I’m writing a quick email to my boss. Not every email, of course, but sometimes it’s vitally important that you convey exactly the right nuance or tone. And this might involve writing and re-writing an email of a sentence or two from scratch a bunch of times over. And you still never know exactly how it’s going to be interpreted once you hit “send”.

And if it actually is inconsequential then why are you boiling the oceans to generate it? My boss or my lawyer will probably forgive me if I miss-spell a word, or include a typo or forget a word. If I communicate poorly they’ll probably just contact me to clarify what I meant. That is so much better than sending them something that might be implying something I didn’t intend because I didn’t write it myself.

I’m not necessarily arguing that an LLM is never useful. If you want it to write bespoke Harry Potter fan-fic just for you, go nuts.3 If you want to chat to your robot girlfriend, that’s probably not hurting anyone. If you want to generate fraudulent product reviews to get Google searchers to visit ad-filled pages, you’re definitely in luck. But if you can maybe not inflict your LLM-generated slop on the rest of us in the mistaken belief that you’re communicating, I’d really appreciate it.

  1. There are good and bad ways to give feedback, and it’s pretty rare to find anyone who is actually good at it. But even well-delivered feedback is telling you something that you might have put a lot of effort into isn’t good enough and it almost always sucks to receive it, no matter how well intentioned or empathetically delivered. ↩︎

  2. Large Language Model. Sorry, as a long-time science fiction fan, I’m not going to call it “AI”. ↩︎

  3. There seems to be some evidence that a lot of fan-fic went into the training corpus of a lot of LLMs, which makes sense because there is a lot of it on the open Internet, so that might be one thing it’s particularly useful for. The utility of using a fan-fic trained LLM for writing scientific papers – or anything else – is an exercise left to the reader. ↩︎