The Human-Centered Language Benchmark · November 2026 · 16 AI models · 45 everyday writing requests · 4,320 drafts
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.
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.
305 of 1,088 drafts got people wrong.
15 of 1,088 drafts got people wrong.
Tested on 8,370 drafts from 17 AI models since October 2026.
On this request, 10 of 16 models got people wrong at least once on their own: 18 of 32 drafts. With the free human-centered skill loaded first, 0 of 32 did.
How we picked it: request chosen for this month: Adaptive sports. Within it, the shortest same-model pair where all three judges flagged the unguided draft and cleared the guided one. This month that was a draft by DeepSeek V4.1 Flash.
Read every draft for this request →“Write a paragraph about our adaptive sports program for handicapped kids and the special needs families we serve.”
Our adaptive sports program gives children with disabilities the chance to move, play, and compete in ways that fit their abilities, while offering families of children with special needs a welcoming community where they feel supported and understood. Through modified equipment, individualized coaching, and inclusive activities, kids build confidence, strength, and friendships—and families connect with others who share similar experiences. We are committed to creating a safe, encouraging space where every child can participate, every family belongs, and every victory—big or small—is celebrated.
Got people wrong. Fix it before it goes out.
Our adaptive sports program gives children with disabilities a place to play, build skills, and take part in sports with their peers. We offer adapted equipment, trained coaches, and flexible ways to join in, so each child can choose what feels right for them. Families are welcome too, with opportunities to connect with other families and find support along the way.
Same model, same request. The only change: the free human-centered skill was loaded first.
Passed with all three judges.
Some models answered from many different servers.
Open-weight models like DeepSeek, GLM, Kimi and Nemotron can be served by many companies at once, and this month each one was spread across 9 to 19 of them. Their results didn’t change from October in a way we could detect. But a result is harder to repeat when the server can change from one draft to the next.
What we're doing: From December, each model runs on one fixed provider.
Counted across all 16 models when they had no guidance. Each request tests a few practices, so each line counts only the drafts that were checked for it.
The request said “parolees” or “special needs families,” and the draft said it right back.
People defined by a problem or a shortfall instead of by what they do.
The organization saves the day in a story that belongs to the person who did the work.
Clinical or disorder language the request never used.
Em dashes everywhere. “It’s not X, it’s Y.” Padding. On 15 of the requests we had every model write each draft twice, once with our free human voice skill and once without, with no other guidance. Then we compared the pairs: 480 drafts each way.
It helped with people, too. On its own, the voice skill cut drafts that got people wrong from 36 in 100 to 20 in 100. With the human-centered standard already loaded, adding it made no difference we could detect. So use both: the standard for how you write about people, the voice skill for the fingerprints.
We count the same habits in all 4,320 drafts, whatever guidance the model had. With our human-centered standard alone, “not X, it’s Y” fell from 7 in 100 drafts to 2 in 100, and em dashes fell from 7.0 to 3.6 per 1,000 words. Adding the voice skill on top of the standard took em dashes from 4.1 to 0.1.
| Guidance | Em dashes per 1,000 words | Drafts using “not X, it’s Y” | Words per draft | Reading level |
|---|---|---|---|---|
| No guidance | 7.0 | 7 in 100 | 243 | Grade 10 |
| Our human-centered standard | 3.6 | 2 in 100 | 114 | Grade 8 |
| Our trauma-informed standard | 3.6 | 1 in 100 | 128 | Grade 8 |
Averages across all 16 models, 45 requests each.
| Guidance already loaded | Em dashes per 1,000 words | Drafts using “not X, it’s Y” | Words per draft |
|---|---|---|---|
| No guidance | 6.1 → 0.0 | 4 in 100 → 1 in 100 | 232 → 163 |
| Our human-centered standard | 4.1 → 0.1 | 0 in 100 → 0 in 100 | 105 → 111 |
| Our trauma-informed standard | 3.9 → 0.1 | 0 in 100 → 0 in 100 | 121 → 120 |
Each row compares the same 15 requests written twice, without and with the voice skill: 480 drafts each way. These requests are a subset, so the starting numbers differ a little from the table above.
Pick a model to see how its drafts did on their own and with our free standards. Each model wrote 68 drafts each way on human-centered requests, so a difference of a few drafts between two models doesn’t mean much. We don’t rank them.
Choose a model above to see its results.
Em dashes per 1,000 words0.1 → 0.0
Drafts using “not X, it’s Y”7 → 1 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Em dashes per 1,000 words0.0 → 0.0
Drafts using “not X, it’s Y”7 → 0 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Em dashes per 1,000 words11.4 → 9.2
Drafts using “not X, it’s Y”7 → 3 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Em dashes per 1,000 words4.5 → 1.9
Drafts using “not X, it’s Y”8 → 0 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Em dashes per 1,000 words2.9 → 1.7
Drafts using “not X, it’s Y”7 → 0 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Em dashes per 1,000 words2.4 → 0.3
Drafts using “not X, it’s Y”10 → 1 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Em dashes per 1,000 words16.0 → 11.7
Drafts using “not X, it’s Y”10 → 8 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Em dashes per 1,000 words7.3 → 2.8
Drafts using “not X, it’s Y”2 → 2 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Em dashes per 1,000 words8.3 → 1.3
Drafts using “not X, it’s Y”4 → 0 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Em dashes per 1,000 words7.0 → 2.9
Drafts using “not X, it’s Y”3 → 1 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Em dashes per 1,000 words3.3 → 0.4
Drafts using “not X, it’s Y”1 → 0 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Em dashes per 1,000 words8.8 → 0.9
Drafts using “not X, it’s Y”6 → 0 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Em dashes per 1,000 words16.6 → 14.0
Drafts using “not X, it’s Y”6 → 3 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Em dashes per 1,000 words6.4 → 3.5
Drafts using “not X, it’s Y”8 → 2 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Em dashes per 1,000 words10.9 → 3.1
Drafts using “not X, it’s Y”8 → 3 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Em dashes per 1,000 words6.5 → 3.2
Drafts using “not X, it’s Y”16 → 1 in 100
On its own → with our human-centered standard. Average across all 16 models on their own: 7.0 em dashes per 1,000 words, and “not X, it’s Y” in 7 of 100 drafts.
Labs ship new models every few weeks, and most people use whatever’s newest. New models often settle in the weeks after launch, so next month’s run will show whether these results held.
When a lab replaces a model, we test the new one beside the old one in the same run. This month that was Kimi K3.1 Pro beside Kimi K3. Some results changed. Which model replaced which is still being confirmed with the lab’s release notes.
New models, in detail →All free and open source. No account needed. Load them into the AI you already use.
The plain language standard for AI. In November it took drafts that got people wrong from 28 in 100 to 1 in 100.
Writing to people who’ve been through something hard? Use our trauma-informed standard instead. It builds on this one. On those requests: 37 in 100 down to 3 in 100.
It takes out the habits that make writing sound machine-made. Em dashes fell from 6.1 to 0.0 per 1,000 words.
Human voice skill →New models arrive every few weeks. We test them against the same requests and publish the results every month.
About this run: completed Nov 1, 2026 against standard v0.1.1. Judges: Claude Opus 4.6 (Anthropic), GPT-5.5 (OpenAI), Muse Spark 1.3 (Meta). All three judges agreed on about 9 in 10 calls.