Using AI for Writing like a Responsible Adult

Using AI for Writing like a Responsible Adult

Plus! Granular Price Discrimination; ETFs; Form Factors; Training Data

1st June 2026

Byrne Hobart

Contents

In this issue:<br>Using AI for Writing like a Responsible Adult—It's incredibly annoying when people spam AI-generated writing everywhere. But it's equally-but-differently annoying to pretend that any use of AI contaminates writing. There are sensible, responsible ways to use AI, but you have to remember—if you're trying to produce original ideas, you have a partly-adversarial relationship with models that are designed to make you feel like that's what you're doing, whether or not it is.<br>Granular Price Discrimination—OpenAI wants to subsidize every line of code that might include an API call to OpenAI<br>ETFs—It's optimal for ETFs to launch before it's clear that there's demand.<br>Form Factors—Meta would love to listen to every conversation on earth and use them for training data. They can't do that, at least not yet, but they can make progress.<br>Training Data—Remember that we're in an era where startups aim to transfer wealth from VCs to consumers—but with the hope that this leads to profits later.

The Diff June 1st 2026<br>0:00/605.857959<br>1×

Talk to this post with ReadHaus.<br>Using AI for Writing like a Responsible Adult

Apologies in advance: publication may be disrupted in the coming weeks. I’ve apparently managed to rupture my patellar tendon. This is one of the important ones, at least if you need to walk. So, expect a few missed posts and some abbreviated ones in the next week or two.

Anyway, onward.

Technology moves faster than norms, and sometimes you end up with a shearing effect where the same thing is simultaneously the subject of delusional promotion from its fans and differently deluded condemnation from its critics. So AI-generated writing is simultaneously liberating, and drivel, and pushing out all of the human stories, and incapable of ever replacing them. There's less of a clear market for the reasonable opinion that almost everybody holds, i.e. that LLMs are a good tool for writing nonfiction, that they're getting better, but that they're also dangerous in subtle ways, and the danger is getting subtler.

Ask for edits; don't ask the LLM to edit : one very defensible use case is that you've written a draft, and you want someone to read it (ideally quickly) before you show it to everyone at once. So you ask the computer. Models don't have great taste in writing, but they do have consistently above-average taste in every possible kind of writing. If you've written a political biography, GPT-5.5 isn't going to give you better feedback than Robert Caro would. But it will give you better feedback than Caro actually will, because he's busy (and please don't interrupt him. He's almost done). It's slightly annoying to have a draft side-by-side with suggestions and to manually type them in; it's much more annoying to realize that one-shotting draft-to-final replaced your favorite line with a contrastive parallelism. There are people who object to even this, but unless they've sworn off Google Docs entirely (or at least turned off its grammar and spellcheck), they're actually still using LLMs to edit their writing all the time.

Autodidacts, or people just getting up to speed in some new space, can flail around a lot because they don't have a good map of common knowledge. They'll reinvent things, misunderstand things, learn concepts but not labels and vice-versa. This is mostly a matter of cumulative exposure to the topic, but LLMs can help you skip a step; they're very good at providing overviews of the literature, recommended places to start, and prerequisites . This is a case where their averageness is a virtue; any given professor might have peculiar opinions on some thinker, which will distort their syllabus. But the average professor's idea of the best way to start approaching some topic, especially if it's qualified with some reference to why someone might reasonably choose an alternative, is actually pretty good guide and roughly what you’d want. (For many programming and adjacent topics, there's a version of it that helps you ship software and a version that could help you prove some original theorem. These are overlapping areas, but usually someone interested in e.g. linear algebra has exactly one of these two use cases in mind.)

They're good at cross-tabulating unstructured data: Back when SEO was a more dominant strategy for getting traffic, a popular format was the top-N list. What publishers like about it is exactly what writers hate about it: the whole idea is to reprocess information that's already out there into some list, and to perhaps add some low-effort snark or attempt to judge it a bit. So, there are a lot of lists out there, both objective ("biggest explosions ever") and subjective ("Columbus' tastiest sandwiches"). One thing LLMs are pretty good at is creating the lists that...