This article exists because I was (maybe rightfully) anxious about publishing my thoughts. There’s no tracking here so I won’t even know if you read it, but if you did, thank you.
If you find yourself doing any kind of knowledge work post-2025 you’ll inevitably have to consume copious amounts of LLM-generated text. Contrary to what the AI tool providers tell us (“uuu our chatbot is so smart and dangerous uuu I’m almost crying”), free-form writing their creations produce is never good, neither in content nor in form. Sure, LLMs have their uses, such as automating the boring parts of coding, which is most of coding, but their performance on anything not done in a quick loop with clearly verifiable goal is still hit and miss. When it comes to free-form writing, it’s mostly miss.
The fact that you have to read more and more of AI slop is not the fault of the models though, because they are just inanimate tools other people choose to use. Why they make their choice dictates they way the plan, roadmap, technical request for comment or just a simple email will be broken:
They were mandated - in this case the author did not really make a choice, someone else up the management chain did. I feel strong empathy for people who lack the power to choose the tools they perform their work with, so let’s not dwell on this case.
They wanted to outsource their thinking to the model - that’s by far the worst motivation. Outsourcing thinking to an LLM means just shifting the thinking burden onto the people who’ll have to review or implement whatever decision the train of “thought” arrives at. LLMs can help with reasoning, but I have yet to see an example of a good decision-supporting artifact (RFC, plan, roadmap) where LLM was in the drivers seat. Maybe the models are too dumb, maybe we suck at composing the relevant context, but regardless of the reason so far it’s impossible to replace humans in the high-level decision making. This is the outsourced category.
They did the thinking but handed over bullet points to the LLM for expansion - Documents born this way might arrive at the correct conclusions, but the reasoning will likely be obscured by insane verbosity, pointless adjectives and overall vibe of the text being long-form LinkedIn post. You have a succinct representation of you thinking in form of bullet points? Great! Save my time, massage the bullet points a bit and give them to me directly! As a non-native speaker (sure you can tell by this point in the text) I’m also sometimes anxious about my English, but I much rather deal with imperfect sentences than with a regurgitation of those sentences into much-too-big bowl of slop. Let’s call this the copywritten category.
Given that the powers-that-be seem to be hopelessly in love with LLMs, LLM-generated text is here to stay in the professional setting. I started to wonder: what are the long-term effects of me consuming this stuff?. Will I start expressing myself or thinking like a Linkedin-dwelling shapeshifter pretending to be human?
I don’t regard “outsourced” category as dangerous in this context, besides the slim risk of facepalm-induced CTE. Once it’s clear that the piece of writing I’m supposed to review is not really embedded in the context of what we want to achieve, or even worse, full of hallucinations, I can safely stop engaging with it, thus protecting my brain.
In contrast, the “copywritten” category is the insidious one. Someone had a good idea, did their research but then decided to use an LLM to regurgitate their points into a taxidermy of human-written text. On one hand, engaging with such artifacts is painful; on the other, it’s not someone requesting me to do their work for them, so I should try my best to help. And I do. I wade through the morass of consultant-sounding drivel, avoid the swarms of pointless adjectives to find the well-reasoned ideas lurking below the surface.
My idea for compensating the mandatory slop consumption is to go back to reading more fiction, the weirder (= farther from the average text) the better. Books like “House of Leaves”, “2666” or weird stories of Lovecraft, Ligotti and Langan (my favorite!). Some of those text even hit two times, the first one when reading them and the second time when actually grasping what the author really wanted to say. That’s a stark contrast with most of LLM text, which strikes only once, when you realize all those sentences are there just to hide the bottomless void where the actual content should be.