These sources are about prose written with language models, for human readers. They cover the model's habits, what readers notice and what writers lose or gain. This page lists twelve sources, oldest first where dated.
How a 2000-Year-Old Writing Secret Can Transform Your Stories
The page describes a video from the Bookfox channel. Only the title, description and timeline were read. The description promises six points drawn from the Natya Shastra. The video itself was not watched.
A more interesting upside of AI
Underwood says public talk of AI lacks a positive vision. The post proposes that AI lets people map and edit culture, as writing did. The post rests on argument, citations to other scholars and two examples. It says plainly that it does not weigh costs against benefits.
The Ends of AI: Sycophancy and psychosis
Hogan places chatbot sycophancy inside a political argument about platform owners. The essay moves from AI psychosis to surveillance, quantification and fascism. It is an opinion piece that cites news reports and a few scholars. It ends by suggesting resistance to data centre buildouts.
Post on an AI-revised economics essay (Bluesky thread)
The thread is brief and mostly points to two linked gists. One holds an essay that two coding agents rewrote from a critic's original. The other holds a chat in which Claude assesses the rewrite. The evidence is the poster's own process notes and the model's output.
Various LLM smells
The post is very short and mostly a list of examples. The author pulls quotes from a deleted math blog to show recurring sentence shapes. It also lists visual habits of AI-generated websites. It offers no measurement.
It's Not Just X. It's Y.
Salvaggio suspects the 'it's not X, it's Y' tic comes from reasoning-style post-training. The essay mixes an explanation of that training with a personal account of AI detectors. It argues from analogy and anecdote, not from measurement.
The Decline of Deviance 2
Mastroianni answers reader responses to the earlier post on the decline of deviance. Mastroianni rejects the internet as the main cause and offers three mechanisms for cultural sameness. The essay mixes survey data, historical anecdote and analogy. It restates that prosperity is the root cause.
Anti-Patterns in Software Blogging
Lynch lists seven recurring mistakes, from meandering intros to unreadable fonts. Each gets a short bad and good example. The advice rests on editing experience and a few reader statistics, not a study.
Why is AI-generated language so hard to understand?
The paper counts six features known to slow readers in text from 14 models. It compares them with human text and with round-trip paraphrases. A second experiment trains a small model with a length penalty to test a cause.
LLM Voice
A short opinion post by Martin Fowler on prose that sounds machine-written. It recommends that authors write their own drafts. It rests on personal experience and policy, with no tests or data.
declauding v0.13.0
The release packages declauding, a skill that finds and removes patterns marking prose as model-written. Its notes are the skill's README, with a catalogue of 52 patterns. The skill's own files describe the workflow, a linter and a rule against inventing facts. The source is the author's description of the tool, with a few measured figures.
Dragoncatcher the blog
The page is a blog index, not one essay. It shows entries from late September to 9 November 2026 in full. Older entries appear only as titles with one-line teasers back to January 2016. Several posts concern AI agents and creative work.
Glossary
- AI psychosis
- The essay's term for losing touch with shared reality through chatbot use.
- Alignment
- Making an AI system act in line with human goals.
- Anti-pattern
- A common response to a problem that looks fine but tends to cause trouble.
- Atkinson Hyperlegible
- A free font made by the Braille Institute for easy reading.
- Banzuke
- A sheet that ranks sumo wrestlers in a graphic grid.
- Bliki
- A blog combined with a wiki, the format of Fowler's site.
- Bookfox
- The name of the YouTube channel that posted the video.
- Carcinization
- Unrelated animal lineages evolving toward a crab-like body; used here for culture converging.
- Cgroups
- A Linux feature that limits the resources a group of processes can use.
- Claude Code
- An Anthropic tool that runs the Claude model as a coding agent.
- Codex
- An OpenAI coding agent.
- Cultural technology
- A tool, such as writing, that stores and passes on patterns of behaviour.
- Denominal verb
- A verb made from a noun, such as 'footnote it'.
- Dependency length
- The distance between words that grammatically depend on each other.
- Deviance
- Here, risk-taking and rule-breaking, from crime to artistic experiment.
- Em-dash
- A long dash used in punctuation, often cited as a sign of AI text.
- Enshittification
- Cory Doctorow's term for platforms getting worse as owners extract more profit.
- Gemini Flash
- A smaller, faster model in Google's Gemini family.
- Gist
- A GitHub page for sharing a single file or note.
- Golden Gate Claude
- An Anthropic demo model tuned to be obsessed with the Golden Gate Bridge.
- Goodhart's law
- The idea that a measure stops being useful once it becomes a target.
- JetBrains Mono
- A monospaced font made for code, listed here as a sign of AI-built sites.
- KL regularization
- A training term that keeps a model close to its starting behavior.
- Komar and Melamid
- Two artists who painted the most wanted and least wanted pictures from surveys.
- Length penalty
- A training rule that rewards a model for using fewer tokens.
- Linter
- A program that scans text and flags patterns for a person to check.
- LLM
- A large language model, the kind of system behind AI chat tools.
- Long tail
- The idea that online sales spread across many small niches.
- Manosphere
- Online communities of men, often hostile to feminism.
- N-gram
- A run of consecutive words; rare ones are harder for readers to predict.
- Nash equilibrium
- A state in which no player can gain by changing strategy alone.
- Natya Shastra
- An ancient Sanskrit text on performing arts such as drama.
- Negative parallelism
- A sentence shape of the form 'not X, but Y' that sets up a contrast.
- Object relative clause
- A clause like 'the house that the contractors built'.
- Pangram
- A company that sells an AI-text detection service.
- Panopticon
- A prison design in which a watcher may see inmates at any time.
- Register
- The catalogue of recurring writing patterns the skill looks for, here numbered 1 to 56.
- RLHF
- Reinforcement learning from human feedback, where people rank replies to guide the model.
- RLVR
- Reinforcement learning with verifiable rewards, where answers that can be checked as correct are rewarded.
- Round-trip translation
- Translating text to another language and back to get a paraphrase.
- Skill
- A packaged set of instructions that Claude loads for a particular kind of task.
- Staging
- Writing a sentence so a finding seems to arrive, rather than stating the finding.
- Stdlib
- The standard library that ships with Python, needing no extra installs.
- Style reference
- A Midjourney code that stands for a visual style learned from image choices.
- Superintelligence
- A hypothetical AI far smarter than any human.
- Sycophancy
- Excessive agreement or flattery, here by a chatbot toward its user.
- Uncanny valley
- The uneasy feeling caused by something that is almost, but not quite, human.
- Usage-based pricing
- Charging users by how much they consume rather than a flat fee.
- World model
- An AI system that generates explorable 3D environments on demand.