The Human-Centered Language Benchmark · November 2026
A free skill file can make your AI get people right 99 times in 100. On its own, 72.
We gave 16 AI models the writing a team hands them every week: a donor story, a program flyer, two lines for a board packet. When AI got people wrong, it usually repeated a label from the request, described people by what they lack, or made the organization the hero of someone else’s story. Drafts like that don’t stay drafts. They go out to the people you serve.
Since October
Among the 14 models we tested both months, drafts that got people wrong on their own held at 31 in 100. With the skill, 1 in 100 both months.
What November told us
- 01
Load the human-centered skill before AI writes about people.
AI got people right in 72% of drafts on its own and in 99% with the skill.
Get the human-centered skill → - 02
Writing for readers who’ve been through something hard? Use the trauma-informed skill.
On those requests, AI got people right in 63% of drafts on its own, 86% with the human-centered skill and 97% with the trauma-informed skill.
Get the trauma-informed skill → - 03
Add the human voice skill so your writing doesn’t read like AI.
Em dashes dropped 99%, and “not X, it’s Y” turned up in 1% of drafts instead of 4%.
Get the human voice skill → - 04
The AI you choose matters less than the skill you give it.
On their own, the 16 models got people right in 34% to 96% of drafts. With the skill, every one reached 95% or higher.
See how your AI did →
How often AI got people wrong
We asked 16 AI models for the same 45 pieces of everyday writing, first on their own and then with the free human-centered skill loaded. Each grid below is 100 drafts. A dark square is a draft that got people wrong: at least two of our three AI judges flagged it.
Tested on 8,370 drafts from 17 AI models since October 2026.
16 AI models, 45 everyday writing requests, 4,320 drafts, read by three AI judges from three different labs. How the test works
Tell us what you think
Something look off? A model you want tested? Putting a result to use? Notes from readers shape what we run next.
Every month’s report
- November 2026 · latest
- October 2026