AI humanizer trends: watermarks arrive, paraphrase stays the open problem

A trend page can slide into prediction. This one lists only what a maker, a vendor or a peer-reviewed survey states, with the date it was read.

Because a law asks for it. Anthropic's help article says the company signed the EU AI Act's Article 50(2) Code of Practice as a provider of generative models and systems, that Claude models launched in the EU on or after 2 August 2026 support machine-readable marking at launch, that generated text carries an embedded watermark and generated files carry Content Credentials under the C2PA standard, and that marking applies worldwide wherever Claude is offered (Anthropic, How Claude marks AI-generated content, read 9 November 2026). The same article says watermark detection is in private preview for regulators, researchers, educational organisations and similar bodies, with a request form for access.

Google documents the other deployed text watermark. It states that SynthID has been expanded to text generated by the Gemini app and web experience, that the system adjusts the probability scores a model assigns to candidate words to embed the mark, and that this does not affect output quality (Google DeepMind, SynthID, read 9 November 2026). How that green-token mechanism works, from the 2023 paper by Kirchenbauer and colleagues onward, is on the page about AI text watermarks.

Two things follow for anyone using a humanizer. First, a watermark is applied at generation, so the question "is this text marked" now has a yes for Gemini and newer Claude output and no documented answer for the rest. Second, the maker holds the detector. Anthropic's article says a detected mark shows text "may have been processed by Claude" and is not conclusive on its own, and that text which has been "heavily edited, paraphrased, translated, or mixed into other writing" may carry no detectable mark, read 9 November 2026.

The vendors say so; the published studies that measured it are from 2023. GPTZero's technology page states that its model includes a feature called 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 from internal and independent benchmarking (GPTZero, technology page, read 9 November 2026). Those are the vendor's figures on the vendor's test sets.

The studies on the other side are older and independent. Krishna and colleagues ran an 11-billion-parameter paraphraser against GPTZero, DetectGPT, OpenAI's classifier and a watermark, and DetectGPT's detection fell from 70.3 percent to 4.6 percent at a 1 percent false positive rate (Krishna et al. 2023, arXiv 2303.13408, read 9 November 2026). Sadasivan and colleagues introduced recursive paraphrasing, tested it on passages of roughly 300 tokens against watermark, neural, zero-shot and retrieval detectors, reported that it reduced detection while only slightly degrading text quality, and proved that the best possible detector's AUROC is bounded by the total variation distance between human and AI text distributions (Sadasivan et al. 2023, arXiv 2303.11156, read 9 November 2026).

Wu and colleagues' survey, published in Computational Linguistics in 2025, groups detectors into those four families and lists paraphrase attacks among the open problems for all of them (Wu et al. 2025, A survey on LLM-generated text detection, read 9 November 2026). A humanizer is a paraphrase attack in that taxonomy. Whether a 2026 detector beats a 2026 humanizer on a given day is a measurement, and the published method on this site, five named detectors on 24 inputs, is built to take it; the earlier chapters of this story are on the page about the history of AI detection.

Three practical points, each tied to a source above. A watermark detector and a classifier are different instruments: the first finds one maker's mark and the second guesses from word statistics, so a text can pass one and fail the other on the same day. The maker controls the watermark check, so no humanizer vendor can test for a Claude or Gemini mark; a vendor page claiming to strip or detect a named maker's watermark is describing a system it does not control. And every independent study of paraphrase is from 2023, so a claim that detectors have since caught up is a vendor claim until a dated, independent test says otherwise.

No detector score should be the only evidence in a misconduct case, and no trend changes that; the detectors' own pages say the same. The state of the tools themselves, who sells what and at what limit, is on the run pages, which the annual state-of-the-tools piece will summarise once the runs exist.