The Voice Drifts Back: AI Slop, Concise Mode, and Simplified Technical English

I have been fighting this for months, and until last week I assumed it was a me problem. Every AI tool I use has a written instruction telling it how to write. Every one of them slides back.

Then forty-eight hours in August turned my private annoyance into a software category.

Two Days

• August 20: vomit hits the front page. It feeds Claude’s answers through a second, smaller AI on your own machine, to “convert Claude’s token vomit into English.”
• August 20: Anthropic ships Claude Code 2.1.237 with a built-in Concise style that “leads with results and skips preamble and narration.”
• August 20: “Don’t paste the AI, please” hits the front page, 1,041 points.
• August 21: Claudette hits the front page. It hands Claude’s last response to Gemini and prints the translation verbatim.
• August 21: “I’m becoming AI-blind” hits the front page, 475 points.

Underneath it all, the year’s highest-scoring Hacker News thread is still the moderators writing “Don’t post generated/AI-edited comments. HN is for conversation between humans.” 4,229 points.

What Is Actually New

Complaining about how models write is not new. Building infrastructure to route around it is.

Neither tool is an instruction. Nobody is asking Claude nicely or adding a rule to a config file. Both hand the finished text to a different AI and have it rewritten.

This skill accepts that no amount of prompting fully cures this, and does the honest thing instead: it hands the response to a different model. Claudette has pinky promised to print the translation verbatim, because letting Claude “tidy up” the translation reintroduces exactly the voice being removed.

— Claudette (nobuzz) README

Now Anthropic’s changelog, version 2.1.234, three days before Concise mode:

Fixed custom, project, and plugin output styles drifting back to the default voice mid-session.

Two independent developers and the vendor’s bug tracker, same week, same behaviour. The voice is not an instruction the model fails to follow. It is where the model goes when nothing is actively holding it elsewhere.

What The Voice Is

Claudette’s documentation describes it better than any research paper: Claude “talks like it’s delivering a TED talk about its own” work. You ask why a test keeps failing and you get a load-bearing assumption, three numbered revelations, and a note that the third is the most instructive.

The tells are hard to unsee once named. The escalating three-part list. “Here’s where it gets interesting.” The dramatic single-sentence paragraph. The kicker. The insistence that nothing is ever just a bug.

Claudette’s before-and-after is worth studying because the after is barely shorter. It is different. The before announces that one small piece of code is the load-bearing assumption of an entire system. The after says that system has three bugs, names the line for each, and states the fix. One is a colleague, one is a keynote.

The escalation is already in people’s config files. One developer described banning not just the tics but individual words the model overuses: “carries”, “hold”, “spells”, “sitting”. A blacklist maintained by hand, one word at a time. Hold that thought.

The Cost Lands On Readers

There’s some psychological mechanism by which my brain immediately recognizes AI generated text and just short-circuits to “there is no information here”. And when I force myself to read AI-generated text I realize I’m making my brain do creative work to impart meaning to the words.

— causal, Hacker News, 21 August 2026

Banner blindness, for prose. Rafal Cymerys, who wrote the AI-blind post, describes design documents with the giveaway phrasing and requirements that read like a machine thinking out loud about what a requirement might be.

The filter does not distinguish text that is generated from text that merely pattern-matches. The penalty lands on anyone whose honest writing resembles the style: every non-native speaker taught to write formally, every technical writer trained on the same register, anyone who read too much of this and picked up the cadence.

The Fix Is From 1983

Not a prompt, and not a second model. A standard written for aircraft mechanics.

ASD-STE100 Simplified Technical English. 53 writing rules and a dictionary of roughly 900 approved words, plus about 1,200 disapproved ones with replacements. Each approved word has one meaning. “Fall” means to move down under gravity. It does not mean to decrease.

European airlines needed maintenance documentation that a mechanic in one country could not misread about an aircraft built in another. Ambiguity there has a body count, so they removed the ability to be ambiguous. One idea per sentence, twenty words maximum. Active voice, name the actor. Simple tenses. No stacked clauses. No hedging.

There are now at least three agent skills implementing it. One, SimpleEnglish, benchmarked six Claude models across eight tasks, 96 runs: a 72.9% average reduction in style violations, and fewer output tokens as a side effect.

“Be concise” is a vibe. The model agrees enthusiastically and writes you a TED talk about concision. “Use one of these 900 words, one idea per sentence, twenty words maximum” is a specification. A violation is a fact rather than an opinion, which is why SimpleEnglish has a number and my old house style rule did not.

What It Costs

• It is ugly for anything that is not instruction. STE was built for procedures. This post is not written in it and would be worse if it were.
• Banning hedging is dangerous with a model. A model that cannot write “probably” will state uncertain things flatly, and confident wrongness beats verbosity for damage. Carve out an exception for uncertainty.
• Short prose is not understood prose. Cutting the theatre does not buy comprehension of the code underneath.
• The 72.9% is the skill author’s own benchmark, measuring conformance rather than whether readers preferred the result.

The BuzzFeed joke in Claudette’s README is a joke. Nobody trained Claude on listicles. The voice is what you get when a model is optimised on human ratings, because a slightly dramatic, structured answer rates well in isolation and is exhausting in bulk.

The Shape Of The Joke

We built the most capable language technology in history, and within two years the working fix for its prose is a controlled vocabulary of nine hundred words, written before the web existed, so someone could service an engine without misunderstanding a sentence.

We did not need better writing. We needed writing that cannot be read two ways, and the aerospace industry solved that in 1983.