Can humanized text be detected? What the studies and the detectors say

Three independent studies from 2023 tested paraphrase against detectors, and all three found it cut detection. None tested a 2026 humanizer.

Krishna and colleagues built an 11-billion-parameter paraphraser, DIPPER, and ran it against GPTZero, DetectGPT, OpenAI's classifier and a watermark. DetectGPT fell from 70.3 percent detection to 4.6 percent at a 1 percent false positive rate; the authors then proposed a retrieval defence, matching candidate text against a store of a model's past outputs, that recovered much of the loss (Krishna et al. 2023, Paraphrasing evades detectors of AI-generated text, arXiv 2303.13408, read 9 November 2026).

Sadasivan and colleagues introduced recursive paraphrasing, applied it to passages of roughly 300 tokens, and reported that it reduced detection across watermark, neural, zero-shot and retrieval detectors while only slightly degrading text quality in many cases. The same paper proved that the best possible detector's AUROC is bounded by the total variation distance between the human and AI text distributions (Sadasivan et al. 2023, Can AI-generated text be reliably detected?, arXiv 2303.11156, read 9 November 2026).

Weber-Wulff and colleagues tested 12 public tools and two commercial systems, Turnitin and PlagiarismCheck, on an original document set. They concluded that the tools were "neither accurate nor reliable", that they leaned toward classifying text as human-written, that overall accuracy stayed below 80 percent, and that content obfuscation techniques "significantly worsen the performance of tools" (Weber-Wulff et al. 2023, Testing of detection tools for AI-generated text, International Journal for Educational Integrity, read 9 November 2026).

That they have caught up. The claims are vendor figures on vendor test sets, which is not a reason to dismiss them, only a reason to date them and wait for an independent test.

GPTZero's technology page states that its model includes a feature it calls Paraphraser Shield, designed to catch AI text altered by paraphrasing or homoglyph substitution, puts its false positive rate under 1 percent, and reports 96.5 percent accuracy on mixed documents (GPTZero, technology page, read 9 November 2026). Pangram's technical report claims error rates over 38 times lower than leading commercial tools on a benchmark of 10 text domains and 8 language models, using a training method it calls hard negative mining with synthetic mirrors (Emi and Spero 2024, Technical report on the Pangram AI-generated text classifier, arXiv 2402.14873, read 9 November 2026). Originality.ai publishes its own accuracy figures and puts the false positive rate of its newest model at 0.5 percent, read 9 November 2026.

The gap between the two tables is the point. The independent numbers are from 2023 and the vendor numbers are from the vendors. Whether a 2026 humanizer beats a 2026 detector is a question nobody outside the vendors has published an answer to, and the difference between a vendor's detector and an independent one is set out on the vendor detectors page.

Because a classifier reads statistics, not history. The detectors this site uses were trained on human and machine text and report how machine-like a passage looks. A humanizer that swaps synonyms but keeps the sentence shapes, the even paragraph lengths and the predictable transitions leaves most of that signal in place. Sadasivan and colleagues' result that detection fell with each recursive pass, rather than at a single switch, is the clearest published picture of that: the signal fades in proportion to how much changed, read 9 November 2026.

Three things decide a reading on a given day:

With named tools and a run ID. The published method runs 24 AI-written inputs and 6 human controls through 12 humanizers, checks every output on GPTZero, Originality.ai, Copyleaks, Pangram and ZeroGPT on the run date, and publishes each reading with a screenshot and the detector's own threshold; the detector columns of the first run will answer whether humanized text can be detected, tool by tool and detector by detector, when it publishes. The method, weights and input set are on the method page, and runs appear under the runs hub as they complete. Each run is dated and never edited after publication, because the only honest answer to this question carries a date.

No detector score should be the only evidence in a misconduct case. The studies above are the reason: the best detectors of 2023 were beaten by a paraphraser, the vendors of 2026 say they have fixed it, and nobody independent has yet published the test that settles it.