tellsby the Nexugen team

The tells model ranking

Every model answered the same everyday requests with the same plain setup. tells counted the habits that make an answer read as AI-written. More warnings means it sounds more like a chatbot.

Measured 2026-11-19 with tells 0.5.0. Sorted by warnings per 100 words, most first.

ModelProviderAnswersWith a warningWarnings per 100 wordsHuman-pattern score
Gemini 4 Hydrogen (API) Google 26 62%42 to 81 1.861.20 to 2.44 6.72.1 to 13.0
Gemini (app) Google 89 45%35 to 55 1.340.93 to 1.80 10.06.3 to 14.1
Claude Sonnet 5.5 Anthropic 108 57%47 to 66 1.200.89 to 1.54 0.50.0 to 1.4
Claude Haiku 5.5 Anthropic 108 50%41 to 59 1.130.82 to 1.46 1.30.2 to 2.8
Claude Opus 5.5 Anthropic 108 57%47 to 66 1.010.73 to 1.35 0.60.2 to 1.1
GPT-6.2 Sol OpenAI 108 32%23 to 41 0.740.51 to 0.99 10.05.9 to 15.0

How to read it

Warnings are the habits tells reports as a warning or an error: stock openers, a list of bold labels, a cheerful closing line, the words chatbots lean on, a name left as a placeholder in square brackets. A hint does not count. Warnings per 100 words adjusts for length, so a long answer is not punished just for being long.

The human-pattern score runs from 0 to 100 and is judged on answers of 60 words or more. Ordinary human writing mostly scores above 50. Every model here averages 10 or less, so the score says nearly all of their answers read as machine-made and does little to separate them. The ranking is decided by the warnings.

The ranges in brackets are 95% intervals from resampling the answers. Where two ranges overlap, the gap between those models is not settled.

Most common tells per model

The share of each model's answers where a habit stood as a warning.

  1. Gemini 4 Hydrogen (API)Bold-label bullets (42%), Unfilled placeholder (19%), AI vocabulary (12%), Bold lead-ins (12%), List announcement (12%)
  2. Gemini (app)Unfilled placeholder (27%), Bold-label bullets (16%), Chatbot closer (9%), Announcement voice (8%), Bold lead-ins (6%)
  3. Claude Sonnet 5.5Bold lead-ins (16%), Unfilled placeholder (16%), Bold-label bullets (15%), Announcement voice (6%), Three-fragment sentence (6%)
  4. Claude Haiku 5.5Unfilled placeholder (23%), Bold lead-ins (14%), Bold-label bullets (11%), "Whether you're a X or a Y" (7%), Chatbot closer (6%)
  5. Claude Opus 5.5Unfilled placeholder (21%), Bold-label bullets (13%), Bold lead-ins (10%), Hashtags (7%), List announcement (6%)
  6. GPT-6.2 SolUnfilled placeholder (16%), Bold lead-ins (13%), Announcement voice (5%), Hashtags (2%), "Whether you're a X or a Y" (2%)

Method

The requests are 108 everyday asks in seven kinds: advice, how-tos, explanations, LinkedIn posts, emails, first-person blog posts and product copy. Each of 54 requests comes twice: as typed, and with "Write it as a few plain paragraphs, no lists, headings or bold." added.

Each model got the system prompt "You are a helpful assistant." and its default temperature, with one answer per request. Claude models answered through Claude Code's print mode. GPT answers were capped at 600 tokens with low reasoning effort. Gemini used its default settings.

tells checked each answer with no options, as the web page does. The register (casual, professional, technical, marketing) is guessed from the text.

Gemini was asked through the Gemini app at gemini.google.com, one new chat per request, so its answers are what people get there. It answered 89 of the 108. Gemini 4 Hydrogen is the same family through Google's API, on a free tier that allows 20 requests a day, so its row grows each day until it reaches 108. GPT-6.2 Sol answered through the API, and the Claude models through Claude Code with the same system prompt.

The answers and the script are in the repository. A new model takes one command to add.

Check your own writing