I agree with the author. Content marketers, journalists and other professional writers are very good at making content sound engaging without losing the "voice" of their clients. Tech writers are generally good at presenting complex topics in very accessible ways. They develop style guides that make this easier to do at scale, and editors who are usually better at reading and tweaking over writing from scratch help put everything together.
AI is certainly able to ingest all of that and create similar work, but doing so requires A LOT of hand-holding. Without that, LLMs bias towards writing like a cross between an exec assistant who just started using Grammarly and a junior beat journalist trying to prove something no matter the framing.
Most people don't do that, mostly because writing is a means to an end for many; thus, I've seen the three-banger punchline style in academic works and Claude's EXTREMELY tired "we have Axios at home" headlines in mediums you'd never usually see them in.
What usually does me in is that LLMs take FOREVER to get their "point" across. Whiz-bang gotcha prose only works if you don't spend five paragraphs telling me something that could have gotten distilled into a two-sentence blurb.
What's worse is that summarizing with AI is really popular but doing so usually leaves out important details, and you never know exactly what unless you skim through the piece. When that piece is AI-generated, as they are increasingly becoming...yeah.
I think this is why I absolutely hate reading AI-generated PR descriptions, Readmes or wiki entries at work. I work with these people every day, we talk to each other in meetings, but then I read perfectly formatted Slop in the freshly opened PR that has your name and picture attached to it, yet it's almost impossible for me to care about it because (subconsciously or not) i can't bring myself to see it as yours. Same or even worse for reviews.
I don’t understand how people aren’t embarrassed putting those kind of PRs online tbh.
Don’t add filler words. Be straight to the point.
Communicating well requires editing. Editing requires going backwards, not just forwards. It requires turning things over in your head. It requires rethinking connections. It requires knowing when one word says more than a sentence. It requires remembering what you were trying to say. It requires remembering.
LLMs don't have much of that. They can mimic some bits with loops, but usually not very well.
I think AI also has a lot of internal conversation - you can see it in the reasoning phase - where it will invent shortforms as it talks to itself. This also makes it really hard to follow AI text because it will often invent its own jargon.
1. It seems we are starting to see more people write like ai, is this something you have measured yet? (I never used em-dash (big fan of the parens), and I resisted for a while, re: ai, but now am trying to use them appropriately)
2. I've been reading from paper again, and my own writing style is evolving, is this too something you can/have measured?
I started using the emdash after learning of the compose key feature of X11. Not tuning the emdash dial to 11 or anything, but occasionally. Now I avoid it to spare myself the indignity of someone confusing me for a chatbot. In general I prefer a more vulgar style of writing as a direct response to those chatbot tics.
Of course, that now distinct LLM style was kind of a thing before, too, but if you weren't writing some bullshit on SEO on Medium or congratulating yourself for your achievements on LinkedIn, using elements of that style wasn't such a source of confusion. Now the chatbots will frame everything as though it's an insight on B2B sales strategies that struck them when they tripped on a dog leash while jogging to work.
I'm very glad of this now, because it means that I can still write with en- and em-dashes, while also demonstrating that my writing is not LLM-generated.
My thoughts on "AI writing" is that the fears around it seem overblown. The frontier models seem to be getting worse, if anything. If one were to make fine-tunes with a specific style, well, that's also a lot of work just to write in a style. How many books in that style can one sell to justify the training? I don't know if the numbers ever add up.
The entire thing might be a great thing for writing and eventually lead to a revival, because now readers will have a motivation to understand the human behind the piece which is always the most interesting part.
Mate we’ve only had AI for like 3 years. “Ever” is the wrong word to use here.
Generating text humans like, empathising and cold reading the user’s intent, observing something interesting if second hand ally the human experience - none of these can be verified without a human. So they rely on larger pre/train and a larger human curated post train corpus. These are not scalable in the same way.
If the ROI was there, they could do more. Startups like sesame and previously pi showed that a more personable agent than today’s frontier models can be achieved with a sane budget. That’s another unverifiable objective.
But I do think this technology that explains the jagged leaps forward this year does not apply to this problem.
Does, "How many p's are in raspberry?" count as a math problem? Because ChatGPT still returns an answer of '2'.
the new version of this is Jev on coin flips and dice rolls
I've been reading an AI-generated technical proposition. It was even slower for me.
> Beyond this, most models are extremely verbose, which makes total reading time much higher.
Painfully true. When I took a semster of law studies and got my first homework assignment, I was flabbergasted by how short the submission length limits were. I could barely write an introduction to my mini-paper with that number of pages! And I was expecteds to cover everything, and have an explicit conclusion, and a complete and meticulous bibliography. I struggled sooo much to be more concise and to let go of less-critical points I wanted to make! But it was a great intellectual and rhetorical exercise.
They were mostly incomprehensible, mutually contradictory, and only tenuously related to the problem we were trying to solve. The poor engineering team had wasted god-only-knows-how-many person-hours building the tool, only for beta users to be completely unable to figure out how to actually use it. Getting things back on track was quite the process. I would estimate that using AI cost the company about 100 times as many person-hours as the original project lead saved themself by avoiding doing any of their own thinking and writing.
In the past two decades of CG etc. untold numbers of words have been spilt on the subject of the Uncanny Valley and you always get this explanation of how the brain has dedicated pattern matching circuitry and can easily flag when faces feel "off".
How interesting that the same is true for language. Turns out we're not just visual-heavy animals, but also very symbolic ones.
AI absolutely has a detectable style once you read enough of it.
(What it feels like to try and get a decent sentence out of my AI.)