Detect AI text and transform it into any voice, from one platform.
Highlighted: reads as AI. Blue: words the rewrite changed.
Everything is starting to sound the same. #
More and more writing passes through a language model before anyone reads it.
The leading models are trained toward one house style: balanced, polished, eager to sum things up. Spread across millions of essays, emails and posts, that style wears away the quirks that let you hear who wrote something. Writing gets more generic, and readers trust it less.
TextStyled works on both sides of that problem. It shows where text reads as machine-written, and it rewrites a draft in a real person's voice, learned from a few things they wrote.
Whose voice is it? #
Each system gets four texts by one writer and a flat draft to rewrite in that writer's voice.
A style model that never saw the rewrites then compares each one with that writer and 59 others. The score is how often it picks the right writer. Chance is about 1 in 60; the writer's own text scores 81.2%.
57.5% of TextStyle v1 rewrites are traced back to the right writer.
Catches machine text, spares human writers. #
On Pangram's own benchmark of 898 answers, half written by people and half by models.
Each detector is held to the same rule: it may wrongly flag at most 1 in 100 human answers. The score is the share of machine-written answers it still catches. TextStyle v1 also passes over 99% of essays by English learners, the writers detectors most often accuse by mistake.
96.7% of machine-written answers caught, with 1% false alarms.
For developers, teams and labs #
API and MCP Coming soon
Detect and rewrite from your own code, or give your AI agents both tools over MCP.
Enterprise
Your brand's voice in every draft, and passage-by-passage AI review, deployed in your own cloud.
For labs
An RL environment for training a specific writing style into your own models.