Every model on our internal writing benchmark was told one simple rule: zero em dashes. They wrote

Every model on our internal writing benchmark was told one simple rule: zero em dashes. They wrote 16,533 of them across 8,050 scripts anyway. Older models were the worst: 6.6 per script for models released before 2026. The 2026 models mostly listen, at 0.5 per script, and GPT-6 Sol and Claude Fable 5.1 used zero. Then I checked my own 10 scripts: 39. Some models listen better than I do haha #short