Do AI humanizers work?

Four studies from 2023 did most of the work the field still relies on. None of them tested a commercial humanizer by name; each tested paraphrasing, which is the operation a humanizer performs.

Krishna and colleagues built DIPPER, an 11 billion parameter paraphraser that can reorder content and control lexical diversity, and reported that it dropped DetectGPT's detection accuracy from 70.3 percent to 4.6 percent at a constant 1 percent false positive rate, "without appreciably modifying the input semantics"; they also proposed a retrieval defence that stores generated text and matches paraphrases back to it (Krishna et al. 2023, arXiv 2303.13408, read 9 November 2026). Sadasivan and colleagues used recursive paraphrasing on passages of roughly 300 tokens and found it significantly reduced detection rates while only slightly degrading text quality in many cases; they also derived a bound tying the best possible detector to how far AI and human text distributions differ (Sadasivan et al. 2023, arXiv 2303.11156, read 9 November 2026).

Weber-Wulff and colleagues tested 12 publicly available detectors plus Turnitin and PlagiarismCheck and concluded that the tools were neither accurate nor reliable, that they leaned toward classifying text as human, and that obfuscation significantly worsened their performance (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026). Liang and colleagues found that several GPT detectors flagged about 61 percent of TOEFL essays by non-native English writers as AI, and that simple prompting strategies both reduced that bias and evaded the detectors (Liang et al. 2023, Patterns, read 9 November 2026). OpenAI retired its own classifier in July 2023 after reporting a 26 percent true positive rate and a 9 percent false positive rate at launch (OpenAI, new AI classifier announcement, read 9 November 2026).

Put together: in 2023, paraphrasing moved detector readings a long way, detectors disagreed with each other, and some flagged human writing. Every one of those findings is dated.

Detector makers have had three years to respond, and three of the five detectors in this site's method publish something about it.

GPTZero's technology page describes a sentence classifier plus a Paraphraser Shield against paraphrasing and homoglyph attacks, and states that its confidence categories are tuned so the average error rate is under 1 percent for high-confidence predictions on an internal evaluation set (GPTZero, how it works, read 9 November 2026). Originality.ai's accuracy page states that its 3.0.2 Turbo model of September 2025 reached up to 97 percent accuracy on the latest AI humanizers and bypassers with a 1.5 percent false positive rate, that the model was retired in September 2026, and that its 5.0 models of October 2026 are stated at 99 percent or more accuracy (Originality.ai, AI detector accuracy, read 9 November 2026). Turnitin states that its indicator covers content modified by AI paraphrasing or bypassing tools and that its false positive rate is below 1 percent for documents with more than 20 percent AI writing (Turnitin, AI writing detection, read 9 November 2026).

All three figures are the vendors' own, measured on the vendors' own test sets. No peer-reviewed study this site could find has reproduced a 2026 detector's humanizer accuracy, which is why the method here records every reading per detector and per date rather than taking a single accuracy claim as settled. Turnitin's detector is not available to the public, so the Turnitin figures on this site come only from Turnitin's statements; the Turnitin evidence page sets them out in full.

The studies above measured detectors. The cost to the reader is a different measurement, and almost nobody publishes it. Krishna and colleagues described their paraphraser's output as not appreciably modified in semantics, by their own automatic measures. Sadasivan and colleagues wrote that quality degraded "only slightly in many cases". Both phrasings are honest about being model-side judgments, and neither tells a student whether a dropped qualifier turned "no significant effect" into "an effect".

Three kinds of damage are the ones to look for in any humanized text:

• Dropped claims: a qualifier, a date or a condition goes missing and the sentence becomes broader than the evidence.
• Invented claims: a near-synonym changes a technical term, or a smoothed sentence asserts something the input did not.
• Introduced errors: grammar breaks, wrong words and sentences that no longer parse, which rise as rewrite strength rises.

The published method on this site rates every output against its input on a 1 to 5 meaning scale by two editors reading blind, counts errors per 100 words, and weights the two together at 55 points of 100 against 30 for detector pass rate and 15 for limits and access. The weighting is the argument of this page in numbers: a rewrite that reads as human and says something else has not worked.

The operation is the same, a model rewrites text; the promise and the controls differ. A paraphraser is sold on wording and tone. A humanizer is sold on how the output reads to detectors, usually with strength modes and a per-request cap on the entry plan. Among the twelve tools in the ranking, Walter Writes AI names Simple, Standard and Enhanced modes with a 750-word cap on its Starter plan, Phrasly names Balanced, Strong and Max, HIX Bypass names Fast, Aggressive and Latest, and QuillBot's humanizer sits inside a paraphrasing suite with Basic and Advanced modes, all checked on 8 November 2026 (Walter Writes AI is one of the twelve).

The strength ladder is the practical difference. A paraphraser has one setting; a humanizer's strongest mode is where the detector reading moves most and, by the logic of the studies above, where meaning is most at risk. The method runs both the default and the strongest documented mode for that reason, so a reader can see the trade.

Three questions sort a vendor claim from a fact. Which detector, on which date, measured the reading? Who checked that the output kept the input's claims? And what plan and request cap did the run use? A claim that answers none of them is marketing. A claim that answers all three is a test, and its figures will move when the detector updates, which is why every reading on this site carries a date.

For the tools themselves, start with the full humanizer ranking and the free humanizer list; for other questions, the guides hub holds the rest of this series, including what Reddit threads recommend and how those picks compare with the vendor facts.