Humanizer benchmark. reright next to 12 other humanizers, on public texts.

We took the AI-written texts and the humanizer outputs that the AI Humanizer Benchmark publishes, ran the same texts through reright, and scored every output with the same open checks. reright ran this test. The code, the data and every output are below, so anyone can check it or run it again.

What we measured

• Facts: whether a number, a name or a quote of three or more words was changed or dropped. We used the checker published with the paper "Humanized but Still True?" (arXiv 2611.00359). We reran it on the benchmark's October outputs first, and it matched the paper on every one of them.
• Meaning: an open model, gpt-oss-20b, compares each output with its source and scores it from 1 to 5. A second open model, roberta-large-mnli, checks how many sentences of the source the output still says.
• AI detection: four open detectors, desklib, fakespot, Binoculars and Fast-DetectGPT. Each one is set to flag 5% of human texts of the same length. A text passes when none of the four flags it.
• Handed back: when one of reright's checks fails, it returns the text unchanged. Those texts count as reright's output in every column.

November 2026

Tools are listed alphabetically. Within a single month, a gap of up to about 20 points between two shares can be chance. The intervals for every comparison are in the full tables.

All three months

This table has the 9 tools that are in all three monthly runs. Over all three months, every other tool changed facts more often than reright, and the 95% interval of each difference excludes zero. reright changed a fact in 1% of the texts that have one: it dropped "first" from "first-time triathletes". HIX Bypass passed the detectors more often than reright. Seven tools kept the meaning less well, with intervals that exclude zero, and one is within the margin. ReHumanize kept the meaning better, and 95% of its outputs failed at least one detector.

Where reright falls short

• It can still drop a word that carries a fact when nothing marks it as one, such as "first" in "first-time triathletes".
• It still hands 4% of texts back unchanged, when it cannot rewrite them without changing a number or a name.
• On the detectors, HIX Bypass passes more often than reright, and 16% of human-written texts also fail at least one of the four.

What this does not show

• Commercial detectors. The benchmark also scores each tool on GPTZero, Originality.ai, Copyleaks, Pangram, ZeroGPT, QuillBot and Grammarly. We have no such scores for reright, so we compare nothing on them.
• How the tools behave today. The competitor outputs are from September to November 2026, and tools and detectors change often.
• Other languages. Every text here is in English.
• Your own voice. These are reright's default rewrites, before it has learned from your writing.

Plans and features

Read from each vendor's public pages on 6 November 2026, without an account. A blank cell means we could not confirm it there.

Data and code

The download has the inputs, every tool's outputs, reright's six runs, every score, the code and the logs. Runs 1 and 2 used the production version of the morning of 6 November 2026, runs 3 and 4 the version of that afternoon, which keeps quotations, greetings and headings as written, and runs 5 and 6 the version of that evening, which also puts back a number or name the rewrite changed. The tables show run 5. The README has the protocol we fixed before reright saw any of these texts, and every change we made to it.

• public-benchmark-2026-11-06.zip
• README.md
• tables.md

Credits

Texts and competitor outputs: AI Humanizer Benchmark, cycles 2026-09 to 2026-11, CC BY 4.0, reformatted by us. The benchmark is run by the team behind UndetectedGPT and has not reviewed or endorsed this test. Fact checker: humanizer-fact-audit, MIT licence. Federal human texts: GoHumanize Open Humanizer dataset, CC BY 4.0. Questions or takedown requests: [email protected].