I don't want to read what you didn't write
TL;DR
Colin Breck argues that AI-written design docs, PR descriptions and messages fail their readers. The reader lacks the context that the sender had when they prompted the AI. Breck uses AI to check facts, add citations and edit his own text. But he found its drafted prose unpleasant and often wrong.
The article
- People increasingly build something with AI and then have AI write the design document after the fact. That turns a tool for building consensus into an exhaustive machine summary nobody wants to read.
- AI-written PR descriptions list every change. But they skip what a reviewer needs to know: why the work matters, how risky it is, and where to look.
- AI output reads well to the sender because the sender holds all the context. The reader has none and has to read every line to reconstruct it.
- A cited survey found that 78% of developers stop reading when they suspect an article is AI-written. 98% prefer the author's own flawed prose.
- Writing a paper, Breck had AI check paragraphs against code, logs and metrics, fill in BibTeX citations, catch grammar errors and draw TikZ diagrams. It caught a notation error four expert reviewers missed. Paragraphs it drafted from his context were unpleasant to read and often wrong. The one exception was the abstract, which he kept unchanged.
- Breck points to ASD-STE100 Simplified Technical English for AI-written manuals. He also points to Pangram, which Oxide now requires public writing to pass as human-authored. He argues that summaries lose the most in narrative and uncertainty.
Why it matters
Worth reading for anyone setting team norms on AI-written docs or PRs. The original has the concrete workflow Breck used on his paper. It also has the long quotes from Simon Sarris, Bryan Cantrill and Iain McGilchrist that carry the essay's case for writing as attention.
The discussion
A huge, mostly sympathetic thread, whose top comment set off a long information-theory argument: can AI add real information to a 300-bit prompt? Dozens of developers swapped stories about bloated AI PRs and specs at work. A loud minority called the essay's stance elitist toward people who struggle to write.
Can AI add information to your prompt?
hatthew argued that if an AI fills in 700 bits around your 300, the reader could infer those 700 too, so send the 300. Critics answered that inference costs the reader effort and that AI knows more than either party. A sender who vouches for checked output adds signal, they said. hatthew's side replied with the data processing inequality, and asked why the reader needs the sender at all if the AI did the thinking.
If it's able to guess those 700 bits correctly, then they aren't true semantic information, and you really only have 300 bits you want to transfer.
In principle, all of mathematics can be inferred from the axioms. The field of mathematics is not superfluous and something a receiver could infer themselves if they wanted it.
Generation is cheap, review is not
Many described the same workplace failure. Colleagues produce pages of PR rationale, specs and ticket comments in seconds, and reviewers must read all of it because approval implies they did. One lead described juniors whose confident AI-written PRs had never been run, and teams where AI reviews AI and nobody really reads. One commenter said verification capacity falls behind generation capacity, which moves the burden from the sender to everyone else.
I don't have time to read it and you put me in the position where I can't afford not to - because approving the PR implies I did and accepted it.
At a broad level, the issue isn’t generated content, it is the ratio of verification capacity to generation capacity (V/G).
Gatekeeping or a fair demand?
gfody, dandare and others said AI lets non-native speakers, people on the spectrum and weak writers be heard, so demanding raw prose is privilege talking. The reply was that AI mangles what those writers mean, so the reader can't tell guesses from intent. fzeroracer, who also finds writing hard, would rather read broken English. A smaller group said only the quality of the text matters, not who produced it.
ai is giving voice to a lot of folks we probably wouldn't even be hearing from otherwise
I would sooner read through an essay with endless punctuation issues and spelling mistakes than one passed through an LLM because one is an earnest attempt to communicate while the other is offloading all of the mental burden to me.
What teams are trying
The most common fix was to ask for the prompt instead of the output, though others countered that a prompt carries less than vetted output. Others use a hand-written "AI;DR" section atop PRs, pair-review huge PRs with the sender, ask coworkers for a non-AI version, or force agents into ASD-STE100. One commenter said a firm "write like a human" instruction to Claude helps. Another tried blacklists and prompt evolution, and nothing removed the Claude voice.
Please don't send me AI-generated text. Send me your prompt instead.
One thing that has markedly helped is tweaking our AGENTS.md prompt and prose-heavy skills like `/pr` or `/ticket` to specify that all technical output needs to be written in ASD-STE100 Simplified Technical English. It has helped us so much.
The "isn't just difficult, it is punishing" line
About a dozen separate threads seized on one early sentence, a not-X-but-Y construction with an em dash, as an AI tell. Some readers said they closed the tab there. Most who looked closer read it as a deliberate joke, and Pangram scored the essay as human-written. Others noted humans used both patterns long before AI, and that people now strip them from their own writing to avoid suspicion.
for better or for worse, those characteristics are just going to set off the AI detector in people's heads now, as evidenced by the fact that there are like 10 separate comments on this post that picked up on the exact same thing.
It's clearly a joke. The rest is readable and well written, and free of LLMisms
Consensus
Unreviewed AI text sent to colleagues pushes the work onto the reader. If the sender didn't read it, nobody should have to. Most agreed that the problem is missing context and intent, more than the use of AI as such.
Dissent
A sizable minority said the stance is elitist toward people who write badly. They added that judging text by its origin instead of its quality is a mistake. And a sender who checks the output and vouches for it does add value.