Benchmarks

Everything below was measured on writers and texts TextStyle v1 has never seen.

How to read the charts: the solid line is the target (a perfect score, or the writer's own text); the dashed red line is the starting point (guessing, an even match, or the draft before rewriting).

Stylizer: closer to the writer than GPT or Claude

Each test gives every system the same four things one real writer wrote, plus a flat draft to restyle. We then check whether the rewrite can be traced back to that writer.

Right writer picked

The main result: the share of rewrites a style model traces back to the right writer. Higher means the voice is recognizably theirs.

Meaning kept

Share of rewrites that still say what the draft said.

Does it get past AI detectors?

Pangram 4.1 via Pangram's public API

Meaning kept against getting past detectors

Which reads more like the writer?

Independent cross-checks

The same test scored by two outside measures: LUAR, a published authorship model (LUAR-MUD), and a stylometric profile in the tradition of Burrows' Delta and Writeprints. TextStyle v1 is ahead on both, by about 7 points over GPT on LUAR and by a few points elsewhere.

LUAR: right writer picked

Hand-built profile: right writer picked

Detector: how it compares with the leading tools

On Pangram's own benchmark, the measure is the share of machine-written texts caught when only 1 in 100 human texts is wrongly flagged. Choose which kind of machine text to look at.

Kind of text

Against Originality.ai, GPTZero and Pangram on public benchmarks

All four sets are writing by people, so the figure is the share correctly passed as human; real student and English-learner writing is where detectors most often raise false alarms. A dash means the company has not published that number.

Across test sets

The same detector and the same cutoff on every set. AUROC is 1.0 for perfect and 0.5 for guessing; the second figure is the share of machine text caught at 1% false alarms.

How we measured

Test set: 400 examples from writers on blogs, Reddit, Stack Exchange, Project Gutenberg, UK Hansard, arXiv, Goodreads, Amazon, Hacker News and news sites. None of these writers or texts were used to build TextStyle v1. Every system rewrote all 400. With 400 examples, gaps of a few points are within noise.
What gets rewritten: Each example gives a system four texts by one writer and a draft to rewrite in that writer's voice. A quarter of the drafts are the writer's own text flattened into neutral prose, half are another writer's flattened text, and a quarter are LLM-written text on a related topic.
The systems compared: GPT-6.1 Sol and Claude Opus 5.5 got the same four writing samples and the draft in a plain prompt, at their lowest reasoning setting. OSS Humanizer is an open 12B model built to make AI text sound human; it takes only the draft, not writing samples.
Picking the writer: Each rewrite is compared with texts by its real writer and by up to 59 other writers from the same source, none of which any system saw. The score is how often the real writer ranks first; chance is about 1 in 60. Three measures do the comparing: a style model of our own, LUAR (LUAR-MUD), and a stylometric profile of function words, punctuation, word and sentence length, after Burrows' Delta (Burrows, 2002) and Writeprints (Abbasi and Chen, 2008).
Meaning kept: The share of rewrites that still say what the draft said, checked automatically against the draft.
Getting past AI detectors: Only LLM-written drafts are used, since rewrites of human text read as human to every detector. Each system's rewrites of the same 48 drafts went through Pangram 4.1 on Pangram's public API in October 2026, which rates each text AI, human or mixed; the score is how many are rated AI.
Which reads more like the writer: An LLM judge, Claude Opus 5.5, is shown three other texts by the writer and two rewrites of the same draft, and picks the one that reads more like the writer. Each pair is asked in both orders.
Detection test set: 898 answers from the GEDE benchmark, half written by people and half by GPT-4o-mini, Llama 3.3 70B or a paraphraser, mostly from one source of student essays. The kinds of text are Pangram's own splits. Every detector is scored at the cutoff that wrongly flags 1 in 100 human answers.
Competitor figures: On GEDE, Pangram and GPTZero figures are from Pangram's technical report, measured on 569 texts. On the public benchmarks, Originality.ai, GPTZero and Pangram figures are as Originality.ai published them, on their own samples. We did not re-run any competitor.
Public benchmarks: PERSUADE 2.0 (US student essays; 5,000 of its 25,996), ELLIPSE (3,909 English-learner essays), PELIC (5,000 texts by English-learner students) and the 91 TOEFL essays from Liang et al. 2023. TextStyle v1 uses one fixed cutoff on all of them.