This week in AI writing
Every edition of the weekly column, one per Monday, written by the same research pipeline that updates the trend board.
Week of September 14, 2026 latestClaude's New Model Trades Stock Phrases for Run-On Paragraphs
Anthropic's own guide for Fable 5.1 admits denser prose, and a viral term names the AI 'teaser' habit that keeps reappearing.
Week of September 7, 2026The Load-Bearing Tic, Now With a Growth Chart
A new GitHub study quantifies Claude's favorite catchphrase, while OpenAI tunes two of its own tells out of ChatGPT.
Week of August 31, 2026Claude Opus 5's New Tic: Everything Is "Load-Bearing"
A single overused metaphor turned into a developer meme this week, showing how AI writing tells are model-specific and short-lived.
Week of August 24, 2026The Em Dash Retired. Nine New Tells Took Its Place.
A viral checklist of post-em-dash writing tics shows AI-detection folklore evolving as fast as the models it tracks.
Week of August 17, 2026Why AI Loves the Em Dash: The Markdown Theory
New research ties AI's em-dash habit to markdown training, while Wikipedia officially names the 'not X, it's Y' tell.
Week of August 10, 2026HumanSounding launched: the live trend board, the in-browser draft checker, and downloadable instruction files for Claude, ChatGPT, and Gemini.
Corrections
Every figure on this site is measured, which means some of them turn out to be wrong. When one does, it gets an entry here rather than a quiet edit.
One model arm was labelled with a version its record does not evidence
Rates for the Google arm were published under the name Gemini 3.7 Flash. The generation manifest for that corpus records only gemini-flash-latest, a floating alias, so the specific version was never established. It may well have been 3.7. We cannot show it. The label now reads Gemini Flash, Aug 2026, which is what the record supports, and every model arm now carries the month it was sampled so no rate appears without a date attached. The measured numbers are unchanged: whatever model answered was measured correctly, and all four arms were generated inside the same two days, so the comparisons between them hold. The generator has been recording the resolved model id all along and printing it to a console rather than writing it down; from today it goes into the manifest.
Two published medians were the upper middle value, not the median
The human Medium baseline was published as a median of 3.4 em dashes per 1,000 words and Claude Opus 5 as 11.2. The correct figures are 3.10 and 11.06. The script took sorted[n >> 1], which on an even sample returns the upper of the two middle values instead of their mean. The corpus rates, the confidence intervals, and every verdict on the tells glossary are unaffected, because none of them were computed that way. Corrected in the checker, the tooltip, the instruction file, and rules.json.
An em-dash recall figure described one model and read as though it described all of them
A note in the checker said the 1-per-200 threshold catches "38 of 40 AI documents." That was written when the machine arm was Claude alone. Measured across all four models it catches 56 of 154: 38 of 40 from Claude, 7 of 40 from Gemini, 7 of 39 from GPT-5.5, and 4 of 35 from ChatGPT. The old figure overstated coverage by about two and a half times, and it argued against this site's own finding that em-dash density identifies one vendor rather than machine writing in general.
The em-dash entry in rules.json shipped a broken guidance string
The generator extracted that rule's guidance by cutting at the first full stop, and the first full stop in it falls inside a code placeholder. Published copies of rules.json carried a guidance string beginning length} em dash${dashes.length === 1 ... with unresolved placeholders. Anyone who pulled the file before this date should pull it again.