How we test AI humanizers

The input set is fixed for a calendar year; a new set starts a new series, and scores are compared only within a series. It has two parts.

The AI-written part is 24 texts across six genres, four per genre: a school essay paragraph set of 350 to 500 words, a blog post section, a product or service description, a professional email, a cover letter, and a research-style abstract. In each genre, two texts come from one widely used chatbot and two from another, generated from the same written prompts. The prompts are published with the run. The generator, model and date sit in the raw file for each input.

The human-written part is 6 control texts, one per genre, written by the editors for this set and never published elsewhere. Their job is to show each detector's false positive behaviour on the same genres, so a reader can see what a human text scores on that detector before judging a humanized one. Liang and colleagues found that detectors flagged 61 percent of TOEFL essays by non-native English writers as AI-generated (Liang et al. 2023, Patterns), which is reason enough to publish a human baseline beside every detector reading.

Every input, AI and human, is checked on every detector before any humanizing. That baseline row makes the change a tool produces visible instead of implied.

The launch set is 12 tools: Walter Writes AI, Undetectable AI, StealthGPT, QuillBot AI Humanizer, Grammarly AI Humanizer, Phrasly, BypassGPT, HIX Bypass, TwainGPT (now Verva), WriteHuman, GPTHuman and Clever AI Humanizer. Each has a review with the same fields, and the main ranking lists all 12 on one table.

A tool is added when it has a public web interface, a stated plan or free tier that accepts a 500-word input, and enough search demand that readers are weighing it. A tool is removed only when its service closes or stops accepting inputs; a renamed tool keeps its row under both names. Changes to the set are dated on this page.

Walter Writes AI is one of the 12 and is scored like the others: same inputs, same detectors, same raters, same fields, including a specific main limitation on its review. It receives no extra placement.

Five detectors, each read on the run date with a screenshot of the reading. "Pass" means the detector's own label is human, or its AI probability sits under the threshold the detector itself publishes; the threshold used is stated per detector in the run notes. Where a detector's page states an error rate, that figure is the vendor's own claim and is recorded as such.

Turnitin is not on the list. Its AI writing detection is sold to institutions and is not available to the public, so this site cannot run it. Turnitin's product page states that its checker identifies when "AI bypassers (also called humanizers)" may have been used and reports likely AI-paraphrased content (Turnitin, AI writing detection page, read 9 November 2026). Pages on this site report that statement and the peer-reviewed literature; no page states a pass or fail result for Turnitin.

Four dimensions, weighted to 100. The weights are fixed for a series and published before the first run in it.

The headline score out of 100 is the weighted sum, to one decimal, with all four sub-scores shown beside it. The formula, the weights and every tool's inputs to it live in one data file that the pages, the structured data and the downloadable CSV all read from, so a number appears once and the same everywhere.

Meaning kept outweighs detector pass rate by design. Krishna and colleagues showed that a paraphraser dropped one detector's detection rate from 70.3 percent to 4.6 percent at a 1 percent false positive rate (Krishna et al. 2023, arXiv 2303.13408); the paper measured evasion, not whether the paraphrased text still made the same claims. Weber-Wulff and colleagues tested 14 detection tools and found all below 80 percent accuracy, falling further on paraphrased and machine-translated text (Weber-Wulff et al. 2023, International Journal for Educational Integrity). Both results show that moving a detector score is easy. Keeping the argument intact while doing it is the harder test, so it carries the larger weight. The guide on whether AI humanizers work covers these studies in more detail.

Price is never scored. There is no "cheapest" or "best value" verdict anywhere on the site; price appears as a fact with its checked-on date.

Each run publishes a page and a set of files, all served from this site's own domain and reusable under the licence in the terms of use (attribution and a link).

Nothing in a published run is edited later. If a vendor disputes a reading, the dispute and the response are recorded in the corrections log, and any change to a number is a new run ID. The editorial policy covers how tools are chosen, how vendor claims are treated and how disputes are handled.

Four facts about what a run can and cannot show.

Detector models change without notice. A reading is true for its date, and the date is on every reading. A tool that passes a detector in one run may not in the next, and the run notes record which detector versions could be identified.

Twenty-four inputs per run is a small sample. The per-detector tables show the spread across inputs rather than hiding it in an average, and the genre column lets a reader see whether a tool handles essays and abstracts differently.

Meaning ratings are human judgment. The rubric, both raters' scores and the third reads are published so that anyone can check a rating against the output text in outputs.csv.

Turnitin is not available to the public. The site relies on Turnitin's published statements and on peer-reviewed studies for anything said about it, and no page states a Turnitin result for any tool.

No detector score should be the only evidence in a misconduct case. GPTZero's own technology page describes its results as a conversation starter rather than a final verdict, and Originality.ai's accuracy page states that detection scores alone should not be used for disciplinary action (both read 9 November 2026). This site publishes the human controls so that readers can see the same thing in the data.

• 2026-11-09: protocol v0.1 published. Input set defined (24 AI inputs across six genres, 6 human controls), launch set of 12 tools named, five detectors named, weights published: meaning kept 35, errors introduced 20, detector pass rate 30, limits and access 15.