WriteVoice
Research summary

How people and AI write differently

We compared people and current AI chat models answering the same prompts, then kept the differences that held up. They became the five WriteVoice dials.

Published · WriteVoice research, 2026

Short answer: when people and AI chat models answer the same prompts, AI tends to write broad, signposted, wrapped-up text: general advice in neat sections, transitions like "first" and "additionally", and a closing summary. People tend to ground claims in specifics, make their point through things that happened to them, and say what they think. On lived experience, taking a side and signposting, the gap between typical AI and typical human answers is about one standard deviation.

Key findings

  1. When people and current AI chat models answer the same prompts, they differ most on five habits: how specific the writing is, whether it draws on lived experience, whether it takes a side, how much it signposts its structure, and whether it ends with a recap.
  2. AI answers tend to be broad, signposted and wrapped up: general advice in neat sections, transitions like "first" and "additionally", and a closing summary.
  3. People tend to ground their claims in specifics (when, where, who, how much), make their point through things that happened to them, and say what they think.
  4. The gaps are large: on lived experience, taking a side and signposting, typical AI answers sit about one standard deviation from typical human answers to the same prompts, and the gap on signposting is wider still for the newest models.
  5. Specificity depends on the model and the setting: most chat models lean general, while some recent Claude models are as specific as people, or more.

From findings to five dials

Each difference that held up became a dial. Your result shows where you sit among people on each one, and where default AI sits.

  • Broad to Grounded: Do you pin things down, or speak in general truths? Most chat models lean Broad; some recent Claude models lean Grounded.
  • Impersonal to Personal: Do your own experiences carry the point? Default AI leans Impersonal: it rarely draws on lived experience.
  • Even-handed to Opinionated: Do you take a side, in a voice that's yours? Default AI leans Even-handed: pros and cons, "it depends".
  • Direct to Signposted: Do you announce your structure, or just say the next thing? Default AI leans Signposted, near the top of the human range.
  • Open-ended to Wrapped-up: Do you stop when you're done, or tie it all back together? Default AI leans Wrapped-up, near the top of the human range.

How we did it, in brief

  • Matched prompts. We compare human and AI writing answering the same questions and prompts, so differences come from the writer, not the topic.
  • Hundreds of prompts, several models. Hundreds of prompts, each answered by people and by several AI models from the Claude and GPT families, in chat-style settings such as answering a question or posting in an online discussion.
  • Discovered, then checked. We looked for the dimensions along which writing varies, kept only those that separate people from AI, and checked them on held-out data that played no part in finding them.
  • Reviewed for stability. Each dial had to hold up across settings and model versions, not depend on topic or length, and mean something a reader can recognise from a short description.
  • Calibrated scoring. The quiz scores your writing with a fixed rubric applied by an AI model, checked against our research scoring on held-out texts, and places you among people answering the same questions.

Limits

  • The dials describe chat-style writing in English. Other genres, such as essays, fiction or formal reports, may differ.
  • "Default AI" means current chat models with default settings. Prompted differently, any model can write more like a person, which is the point of training your double.
  • The dials describe habits, not quality. Being far from AI is not better writing, and being close to it is not proof a text was AI-written.
  • Models change between versions, so we re-check where default AI sits as new models ship.

How to cite

WriteVoice research (2026). The research behind the dials: how people and AI write differently. https://writevoice.ai/research

This page is a summary. The underlying data and scoring are proprietary and not published. Researchers and journalists can ask us for more detail about the methods.

Frequently asked questions

Do AI models write differently from people?

Yes. In WriteVoice research comparing people and current AI chat models answering the same prompts, AI answers tend to be broad, signposted and wrapped up, while people ground their claims in specifics, draw on their own experience and take a side. On several of these habits the gap is about one standard deviation.

What are the biggest differences between human and AI writing?

Five habits: specificity (when, where, who, how much), lived experience used as evidence, taking a clear side, signposting structure with transitions and lists, and ending with a recap. These are the five WriteVoice dials.

Can these findings tell whether a text was written by AI?

No. They describe tendencies across many texts. Plenty of people signpost and summarise, and AI can be prompted to write specifically and personally, so no single text can be judged from them.

Can I get the data or more detail?

We publish a summary rather than the underlying data, which is proprietary. Researchers and journalists can contact us for more detail about the methods.

Now measure your own writing

The same idea, for one person: your five dials from your own words, a voice type, and an AI double trained to write like you.