I'd like to apologize the Byrne for lifting so much of this blog post, but I need to remember this stuff. 

The Diff: AI's Impact on the Written Word is Vastly Overstated, 2025-Mar-3 by Byrne Hobart

If you look at the media environment once printing was established, it actually feels quite familiar: Daniel Defoe of Robinson Crusoe fame was an avid political pamphleteer, including a piece suggesting the mass execution of religious dissenters in England execution (the scholarly consensus is that he's doing what, in modern parlance, we'd call a bit. Like many people on Twitter today, he posted it anonymously, got outed, and found himself in big trouble as a result.) He also wrote ghost stories, foreign policy tracts, opinionated pieces on the money supply, etc. He was, in other words, what you'd refer to as a "blogger" in the 2000s and a "Substacker" more recently: someone with few or no institutional affiliations, who publishes opinionated pieces on a wide range of subjects. …

Pajama-clad or not, bloggers stuck around, and they forced other news organizations to adapt to their norms. Now, it's common for stories to be posted quickly and then repeatedly edited; traffic gets measured and talked about even if it's not a journalist's only performance indicator; major media companies make fewer pretensions to partisanship as a result of competing against independent writers who made no attempt whatsoever to disguise their biases; and blogging melded with the establishment so seamlessly that there are writers today who broke into the lucrative and important business of mainstream media by blogging (e.g. Matt Yglesias) and bloggers who used their mainstream media fame to break into the lucrative business of independent publishing (e.g. Matt Yglesias). In retrospect, the logistics of printing a couple million copies of the same set of text and delivering it by truck should not have determined the entire structure of the prestige news business, but they did, until they didn't…. 

…at this point LLMs make it so that everyone who wants to can write whatever email they need to in whatever tone they want (yes, you can use ChatGPT to Chad up your work emails by removing all of the apologies and clarifications). This can move writing a little closer to 90th percentile, when it's being used by someone who could produce the same result if they were a bit more diligent. But it can also elevate basically any coherent thought into a message that reads like it was produced by a college-educated professional. … It's a huge deal if you can't quickly and superficially judge a text based on those signifiers, and instead have to engage with the ideas before dismissing them as worthless. … 

We already implicitly opt out of the overwhelming majority of what we could read. And whether we read .01% or .001% of what theoretically interests us doesn't make much of a practical difference. After all, we can only consciously process 10 bits of information per second, which amounts to ~2gb in an average lifetime, so we have to opt in wisely…. 

Byrne's Footnote: The AI content of The Diff probably hovers at around the 0.1% mark for text, though AI's share of research is a lot higher and still rising. That 0.1% does not consist of any ChatGPT-generated text; it's actually because one of the steps in the writing process is dumping the text into Google Docs, because Docs is a shared editor and has such good contextual spelling and usage checking, which has been powered by a deep learning model since at least 2021. (In a reflection of just how fast things have advanced: the model has 680m parameters and is described as "a very large model.")