Humanizers versus detectors: the arms race
Humanizers are paraphrasers sold on one promise: the output will pass AI detectors. Since 2023 the two sides have taken turns. Paraphrasing broke the first detectors. Detectors retrained. Humanizers tuned against specific detectors. Turnitin launched bypasser detection on August 27, 2025 and Pangram began training on humanized text in early 2025, both on the theory that rewriting tools have fingerprints of their own. Rewriters retrained again. The result is that no detector score and no humanizer claim from last year says anything about this year, and the market's reviews are written by affiliates because Google will not sell the vendors ads.
2023: paraphrasing wins
The first commercial detectors appeared within weeks of ChatGPT, scoring text on perplexity and burstiness. Within months, two papers showed how brittle they were. Krishna and colleagues built DIPPER, a paraphraser trained to rewrite paragraphs with controllable changes, and used it to push detection rates toward chance. Sadasivan and colleagues showed that running a paraphraser more than once beat watermarks as well as classifiers, and argued from first principles that detection must get harder as models improve. The humanizer market grew out of those results: a web form, a paraphrasing model, and a detector check to show the before and after.
2024: the commodity market
By 2024 the category had a leader, Undetectable.ai, and a long tail of lookalikes priced by monthly word quota. Because Google's advertising policy on enabling dishonest behaviour bars ads for academic-cheating services, none of them could buy search traffic. They competed on organic rankings and affiliate programmes instead, which is why searches for the best detector or the best humanizer return review sites earning commissions on the tools they rank. Independent, dated testing was rare and remains rare.
2025: detectors train on the humanizers
The counter-move was obvious in hindsight. A humanizer is a model, and models have styles. Pangram announced in January 2025 that detectors could be trained to spot humanized text even when the humanizer had been tuned against a specific model, and later claimed detection above 90 percent on high-quality humanized content. Turnitin followed on August 27, 2025 with bypasser detection, trained on humanizer output and counted toward the AI total, English only. Early independent tests, including one by a Northumbria University academic, found the feature caught some tools and missed others, and the vendors on both sides have retrained since.
2026: watermarks enter, and the law
The provider watermarks launched in 2026 do not change the arms race much, because a paraphrase removes them; the providers say so. They change who is responsible. The EU's Code of Practice asks providers to keep marks non-removable as far as they can and to make detection available to appropriate parties. California's platform duty from January 1, 2027 bars large platforms from knowingly stripping provenance. The frontier providers' usage policies restrict building tools that pass generated text off as human, which closes the easiest route for humanizer vendors, who then build on open-weight models instead. Meanwhile the evidence that survives rewriting, document history and provider records, matters more every year.
What a buyer should know
A humanizer's claim is a snapshot against named detectors on an unstated date. Detectors retrain, often without announcement; Turnitin's release notes log changes that marketing never mentions. Academic policies treat rewritten model output as the student's own work only if it is. Process evidence is unaffected by rewriting, and its absence is itself a signal. The humanizer list is maintained for completeness with those notes attached; nothing on this site tests or ranks evasion.
Questions people ask
Do AI humanizers still work in 2026?
Against some detectors sometimes. Turnitin and Pangram now train on humanizer output, and vendors rarely publish dated results, so any claim is a snapshot.
Why are humanizer reviews so unreliable?
Google will not run ads for academic-cheating services, so the market competes through affiliate review sites that earn commissions on sign-ups.
Is it against the rules to build a humanizer on Claude or OpenAI's API?
Both providers' usage policies restrict deceptive uses, including passing generated text off as human work.