Is it safe to use an AI humanizer? What each vendor's privacy page says about your text

Twelve vendors, two questions each: how long is text or personal information kept, and is submitted text used to train models.

Three pages speak to training. WriteHuman's is the most explicit: it states that text submitted to its humanizer and detector, the outputs, and user ratings are used to train and improve its models, that "There is no setting to opt out of this use", and that "Content used to train a model cannot be removed from it", while text sent through its developer API "is not saved to our databases and is not used to train our models" (WriteHuman, privacy notice, read 9 November 2026). GPTHuman states that it "may use Content you provide us to improve our Services", names model training as an example, and offers an opt-out in account settings (GPTHuman, privacy policy, read 9 November 2026). Grammarly's privacy link resolves to the Superhuman policy, which states "We also use information we collect to train our AI models" and offers a training control in account settings (Superhuman, privacy policy, read 9 November 2026).

The other nine say nothing about training. That is not a promise not to; it is an absence. Undetectable AI's and BypassGPT's pages carry the same general retention clause and a policy date more than three years old (Undetectable AI, privacy and BypassGPT, privacy policy, read 9 November 2026). Verva states that user content "may be stored within a user's account history", where users can review and delete it (Verva, privacy policy, read 9 November 2026). HIX Bypass's page covers personal information, usage data and cookies and does not mention the text you submit at all, read 9 November 2026. Walter Writes AI's page makes one statement that bears on where text goes, that it "processes your text using our own self-hosted AI models", and none about whether that text is kept or trained on; its main limitation here is the same as most of the field, a retention clause written for account data rather than for the documents pasted in (Walter Writes AI, privacy policy, read 9 November 2026).

The answer the pages support is: not without reading them, and on some tools not at all for text you cannot afford to have kept. The risk is not the rewrite. It is that a pasted document becomes a stored document, that stored documents can be read by staff or contractors where the policy allows it, and that on at least one tool a document used in training cannot be pulled back.

Two practical rules follow. Strip names, figures and anything identifying before pasting, and paste the sensitive sentence around rather than through the tool. And where a vendor offers an API route with a stated no-training clause, as WriteHuman does, the API is the safer door for confidential work, a point covered on the humanizer API guide.

Sometimes, and in one jurisdiction by statute. Article 50 of the EU AI Act requires deployers of an AI system that generates or manipulates text "which is published with the purpose of informing the public on matters of public interest" to disclose that the text has been artificially generated or manipulated, with an exception where the content "has undergone a process of human review or editorial control" and a person holds editorial responsibility (Regulation (EU) 2024/1689, Article 50(4), EUR-Lex, read 9 November 2026). The same article, in paragraph 2, requires providers of systems that generate text to mark outputs as artificially generated in a machine-readable form, with an exception for systems that "do not substantially alter the input data" (Article 50(2), read 9 November 2026).

A humanizer is a system that alters text, and a humanized article published to inform the public on a public-interest matter falls within the duty unless a human editor reviews it and takes responsibility. For a student, the duty is the institution's policy; for an employee, the employer's. Check your institution's AI policy before using any writing tool. A humanizer does not make undisclosed AI use acceptable, and this site does not help anyone misrepresent authorship.

It does not tell you what the tool does to your text, which is the question the published method on this site measures: claims kept, errors added, detector readings on five named detectors. It does not tell you whether a stated policy is followed. And a policy dated 2023 on a product that has changed every quarter since is a statement about the company's housekeeping as much as its data handling. Read the date first, then the training clause, then the retention clause, and paste last. No detector score should be the only evidence in a misconduct case, and no privacy page should be the only reason to trust a tool with text you could not afford to lose.