Week one: what the measurements say

HumanSounding went live this week. The trend board tracks the words and habits that mark text as AI-written, with every figure traced to a named study. The checker highlights those habits in your own drafts, entirely in your browser. And the instruction files teach ChatGPT, Claude, or Gemini to knock the habits off at the source. Each Monday, this column will tell you what changed.

The first research run already added a tell worth knowing: missing punctuation. Newer models lean toward longer sentences with fewer commas, chaining clauses with "and" instead. The old advice was to watch for em dashes. The new advice is to watch for a paragraph that reads like one long breath.

Elsewhere in the record: Wikipedia's editor guide now flags vocabulary specific to Grok, which reaches for pseudo-scientific words like "causal", "empirical", and "correlate" more than any other chatbot. EQ-Bench now weights the "not just X, but Y" pattern at a quarter of its slop score. And a July analysis found that among the major chatbots only Claude still uses em dashes above the professional human rate; GPT-5.4 now sits below it.

We held ourselves to the same standard. An audit with Impeccable, a detector for AI design tells, found 70 issues at launch, including 42 em dashes on the site that lectures about em dashes. All of them are gone; the whole site now passes its own checker. If your drafts have picked up the accent, the checker will show you where, and the fix takes about two minutes.