stale · formatting · first recorded 2026-07-02 · registers: hn, cooking, parenting
The internet's favorite tell — “the ChatGPT dash” — and our headline negative result: acclaimed pre-AI human prose uses em dashes as much or more than current models do (0.67 vs 0.56 per 1k chars in our measurements). Accuse someone of being AI over an em dash and you will flag more good writers than models.
Kept public as history and as a caution: single formatting marks are terrible evidence. Density and co-occurrence are what count.
model specificity: still ~4× ordinary human web-comment rates, and rising across model generations — but indistinguishable from skilled human writing
⚠ also common in good human writing (collateral score 0.89 — the closer to 1, the more it also fires on acclaimed pre-AI prose)
occurrences per 10k words, measured 2026-07 (run #2: 3 current + 3 vintage models, 3 matched registers, 1,392 samples)
| current AI | vintage AI (2024/25) | humans (matched) | canon (great pre-AI prose) | |
|---|---|---|---|---|
| pooled | 37.99 | 24.50 | 1.45 | 33.73 |
| hn | 32.14 | — | 1.19 | |
| cooking | 27.74 | — | 0.54 | |
| parenting | 35.12 | — | 2.40 | |
| 93.46 | — | — |
regex: —