AI humanizer for journalists: newsroom policies, provenance and what the tools keep
Eight tools, ordered by what the privacy page states about submitted text, from a statement that bears on where text goes to no page at all. Ties are ordered by request cap on the cheapest plan, largest first.
What do newsroom AI policies allow?
Two published policies set the frame, and neither mentions humanizers because both settle the question upstream.
| Policy | Dated | What it states (vendor's wording, read 9 November 2026) | What that means for a humanizer |
|---|---|---|---|
| AP, Standards around generative AI | 15 August 2023 | Staff "do not use it to create publishable content"; "Any output from a generative AI tool should be treated as unvetted source material"; "We urge staff to not put confidential or sensitive information into AI tools" | Humanized copy is unvetted source material until a journalist has checked it; no confidential text goes in |
| The Guardian, approach to generative AI | June 2023, updated March 2026 | Significant AI-generated elements only with "clear evidence of a specific benefit, human oversight, and the explicit permission of a senior editor"; "We will be open with our readers when we do this" | A humanizer pass on a published piece is an AI-generated element that needs permission and disclosure |
AP's statement adds that journalists remain "accountable for the accuracy and fairness of the information we share" (AP, Standards around generative AI, read 9 November 2026). The Guardian's principles, originally published in June 2023, were updated in March 2026 to incorporate guidance following an amendment to its editorial code (The Guardian, approach to generative AI, read 9 November 2026). Read together, the two policies make a humanizer a tool for a draft, used with permission, disclosed when it shaped what readers see, and never fed anything a source would not want stored on a vendor's server.
What do Google and the EU AI Act say about AI-shaped news copy?
Google's position has two parts and neither is about detectors. Its February 2023 guidance states that appropriate use of AI is not against its guidelines and that content produced primarily to manipulate rankings is spam whatever produced it (Google Search Central blog, guidance on AI-generated content, February 2023, read 9 November 2026). Its helpful content page, last updated 5 October 2026, adds that "using generative AI to produce large amounts of text without manual oversight or curation represents little to no effort", that fabricating creator profiles with made-up names or false credentials "is a form of deception", and that AI disclosures "are useful for content where someone might think 'How was this created?'" (Google Search Central, Creating helpful, reliable, people-first content, read 9 November 2026).
| Rule | Who it binds | Disclosure | Where a humanizer sits |
|---|---|---|---|
| Google helpful content guidance, updated 5 October 2026 | Publishers who want to rank | Recommended where a reader might ask how the content was made | A rewrite pass is a "how"; a byline on humanized copy must still be a real person's |
| EU AI Act, Article 50(4) | Deployers publishing AI-generated or manipulated text "to inform the public on matters of public interest" | Required, unless the text "has undergone a process of human review or editorial control" and a person holds editorial responsibility | A humanizer manipulates text; an editor's review and responsibility is the exemption |
| AP and Guardian policies | Their own staff | AP: AI output is unvetted; Guardian: openness with readers | Permission before, disclosure after |
Article 50(4) is quoted from the regulation as published (Regulation (EU) 2024/1689, Article 50, read 9 November 2026 and linked in full on the privacy guide). The practical reading for a newsroom: a humanizer pass on a public-interest story is covered unless an editor reviews it and takes responsibility, which every newsroom code already requires. How to word the disclosure, with MLA and APA forms for the written statement, is on the how to disclose AI use guide.
What does C2PA have to do with humanized text?
C2PA's Content Credentials specification is the provenance standard most newsrooms and camera makers have adopted for images and video. As read on 9 November 2026, the specification site lists versions 1.0 through 2.4 and defines a manifest of cryptographically signed assertions bound to an asset, including an Actions assertion that records what was done to the asset and an Ingredient assertion that records what it was made from, with a trust model built on signer identity and certificate lists (C2PA, Content Credentials technical specification, read 9 November 2026).
For text the standard is early, and this page makes no claim about adoption. The point for a journalist is conceptual: a provenance manifest is a record of every edit, and a humanizer pass is an edit. A newsroom that signs its assets will have to decide whether a rewrite by a third-party model is an action it records, and the vendor pages above give no way to record what the model did. That is one more reason the policies put the human editor, not the tool, at the end of the chain.
One of the 8 humanizers on this page, Undetectable AI, offers a "journalist" readability setting, read 9 November 2026. Three of the 8 state on their privacy pages that submitted text may be used to train their models. That second number is the one a reporter needs first. An AI humanizer for journalists is a tool that rewrites copy for register and for how detectors read it, and it sees every name, quote and unpublished fact in the paragraph it is given. This page sets the newsroom policies that govern the question, then orders 8 tools by what their privacy pages say about text, with the register controls, cap and price beside each. Publishers' concerns about search and provenance are on the AI humanizer for publishers page; the other audiences are in the best-for hub.
Updated
Short answer
One of the 8 humanizers on this page, Undetectable AI, offers a "journalist" readability setting, read 9 November 2026, and 3 of the 8 state on their privacy pages that submitted text may train their models. AP's generative AI standards, dated 15 August 2023 and read 9 November 2026, say staff do not use ChatGPT to create publishable content and treat any AI output as unvetted source material. The Guardian's principles, updated March 2026, require a senior editor's explicit permission for significant AI-generated elements and openness with readers. An AI humanizer for journalists is therefore judged on the newsroom's policy first and on the tool's privacy page second; this page orders 8 tools by what their privacy pages state about keeping and training on text, and scores none of them.
Which AI humanizer for journalists states what it keeps?
Privacy wording is quoted from each vendor's page as read on 9 November 2026 and set out in full, with links, on the AI humanizer privacy guide. Price is reported and never scored. Detector readings and meaning-kept ratings under the published method join every row when the first run publishes.
| Order | Tool | Best for | Main limitation | Free tier | Cheapest plan (checked 9 November 2026 unless dated) | Privacy page on text (vendor's wording, read 9 November 2026) | Register or mode controls | Review |
|---|---|---|---|---|---|---|---|---|
| 01 |  Walter Writes AI | Copy a desk rewrites in 750-word pieces with a strength setting | 750 words a request on Starter; no statement about whether text is kept or trained on | 300-word trial, no card (checked 8 November 2026) | Starter, 30,000 words a month, $8 a month billed $96 yearly | "processes your text using our own self-hosted AI models"; no training statement | Simple, Standard, Enhanced | Review of Walter Writes AI |
| 02 |  StealthGPT | High volume in 1,000-word pieces, with a stated retention limit | No free tier; $1.00 trial rolls into the plan; no training statement | None; $1.00 for 3 days | Pro, 50,000 words a month, $29.99 monthly or $299.99 a year | Kept no longer than "three (3) months past the termination of the user's account"; no training statement | None named on the pricing page | Review of StealthGPT |
| 03 |  Undetectable AI | A register setting that names journalism | Privacy policy dated 18 March 2023 with a general retention clause; no training statement; no request cap stated | 250 words in total (checked 8 November 2026) | 10,000 words a month, $5.00 a month billed $60.00 yearly or $9.99 monthly | Kept "for as long as it is necessary for the purposes set out"; no training statement | Readability levels including journalist, university, high school; purposes including article and story | Review of Undetectable AI |
| 04 | Verva (formerly TwainGPT) | Light use with a detector in the same login | Content stored in account history; no training statement; no free tier; no cap stated | None stated | Basic, 8,000 words a month, $8 a month billed annually | User content "may be stored within a user's account history"; no training statement | Not stated | |
| 05 |  GPTHuman | 1,000-word pieces with a training opt-out | States it may use content to improve services; opt-out is in account settings, not default | 300 words once | Starter, 25,000 words a month, $8.25 a month billed $99 yearly or $15.00 monthly | "we may use Content you provide us to improve our Services"; opt-out in settings | None named | Review of GPTHuman |
| 06 |  Grammarly AI Humanizer | A house voice built from a writing sample, free | Trains on collected information by default with a control in settings; no word allowance published | Free tool and free account; 100 AI prompts a month | Pro, $12, billing period not stated on the card; 2,000 AI prompts a month | "We also use information we collect to train our AI models"; a training control in account settings | 4 preset voices plus a custom voice from a sample | Review of Grammarly AI Humanizer |
| 07 |  WriteHuman | Web users who accept training; API users who do not | States no opt-out from training on web submissions; a request is a request, 600 words on Basic | 3 requests a month, 250 words each | Basic, 80 requests a month, $12 a month billed $144 yearly or $20 monthly | "There is no setting to opt out of this use"; API text "is not used to train our models" | Basic and Enhanced quality; 2 to 5 variations | Review of WriteHuman AI |
| 08 | Clever AI Humanizer | A 3,000-word request free, no account | No privacy page, no terms, no company name visible on the site | 3,000 words a request, no cap stated, no sign-up (checked 8 November 2026) | None; free only | No privacy page visible from the vendor site | Normal, Clear & Structured | Review of Clever AI Humanizer |
Which humanizer should a journalist choose?
One line per job. Each name links to its full review.
- Choose Walter Writes AI if you want a statement that text stays on the vendor's own models and can live with 750-word requests and no retention figure.
- Choose StealthGPT if you want a stated retention limit and will work in 1,000-word pieces at the highest monthly entry price here.
- Choose Undetectable AI if a journalist register setting matters and you accept a 2023 privacy policy that says nothing about training.
- Choose Verva if you want a detector bundled and will delete history yourself.
- Choose GPTHuman if you will switch the training opt-out off before pasting anything.
- Choose Grammarly if a house voice built from your own copy is the point and you will set the training control first.
- Choose WriteHuman only through its API, whose text the vendor states is not used for training.
- Choose Clever AI Humanizer only for text that is already public, because nothing published tells you what happens to it.
How the ranking works
- Inputs
- 24 texts24 AI-written inputs across six genres plus 6 human-written controls, the same set for every tool.
- Detectors
- 5 namedGPTZero, Originality.ai, Copyleaks, Pangram and ZeroGPT.
- Weights
- Meaning kept35
- Errors introduced20
- Detector pass rate30
- Limits and access15
- Run date
The eight tools on the same fields
Every free tier gets the same fields, in rank order. The disc is the sample number it was poured under in the full run.
- Best for
- copy rewritten in 750-word pieces with a strength setting.
- Main limitation
- 750 words a request on Starter, and no statement on whether text is kept or trained on.
- Free tier
- 300-word trial, no card, checked on 8 November 2026.
- Cheapest plan
- Starter, 30,000 words a month, $8 a month billed $96 yearly, checked on 8 November 2026.
- Privacy page on text
- self-hosted models; no training statement.
- Register controls
- Simple, Standard, Enhanced.
- Best for
- high volume in 1,000-word pieces with a stated retention limit.
- Main limitation
- no free tier, a $1.00 trial that rolls into the plan, and the highest monthly entry price on this page.
- Free tier
- none; $1.00 for 3 days.
- Cheapest plan
- Pro, 50,000 words a month, $29.99 monthly or $299.99 a year, read 9 November 2026.
- Privacy page on text
- kept no longer than three months past termination; no training statement.
- Register controls
- none named on the pricing page.
- Best for
- a register setting that names journalism.
- Main limitation
- a privacy policy dated 18 March 2023 with a general retention clause and no training statement; no request cap on the pricing page.
- Free tier
- 250 words in total, checked on 8 November 2026.
- Cheapest plan
- 10,000 words a month, $5.00 a month billed $60.00 yearly or $9.99 monthly, read 9 November 2026.
- Privacy page on text
- kept as long as necessary for the stated purposes; no training statement.
- Register controls
- readability levels including journalist; purposes including article and story.
Verva (formerly TwainGPT)
- Best for
- light use with a detector bundled.
- Main limitation
- no free tier, no request cap stated, and no statement on training.
- Free tier
- none stated.
- Cheapest plan
- Basic, 8,000 humanizer words a month, $8 a month billed annually, read 9 November 2026.
- Privacy page on text
- stored in account history; no training statement.
- Register controls
- not stated.
- Best for
- 1,000-word pieces with a training opt-out.
- Main limitation
- training is the default until the opt-out is set, and the trial is 300 words once.
- Free tier
- 300 words once.
- Cheapest plan
- Starter, 25,000 words a month, $8.25 a month billed $99 yearly or $15.00 monthly, read 9 November 2026.
- Privacy page on text
- may train; opt-out in settings.
- Register controls
- none named.
Sample 06: Grammarly AI Humanizer
Grammarly's humanizer page states the tool is free to use and not intended to change detector readings, and its privacy policy states it trains AI models on collected information with a control in settings, read 9 November 2026.
- Best for
- a house voice built from a sample of the desk's own copy, free.
- Main limitation
- training is on by default with a control to turn it off, and no word allowance is published; plans are metered in AI prompts.
- Free tier
- free tool and free account, 100 AI prompts a month.
- Cheapest plan
- Pro, $12, billing period not stated on the card, 2,000 AI prompts a month, read 9 November 2026.
- Privacy page on text
- trains AI models; training control in settings.
- Register controls
- 4 preset voices plus a custom voice.
Sample 07: WriteHuman
WriteHuman's privacy notice, updated 21 September 2026, states that web submissions train its models with no opt-out and that API text is not used for training, read 9 November 2026.
- Best for
- API users, whose text the vendor states is not used to train.
- Main limitation
- no opt-out from training for web submissions, and a 600-word request on Basic costs the same whether it is 100 words or 600.
- Free tier
- 3 requests a month, 250 words each.
- Cheapest plan
- Basic, 80 requests a month, $12 a month billed $144 yearly or $20 monthly, read 9 November 2026.
- Privacy page on text
- trains, no opt-out on the web; API text not trained on.
- Register controls
- Basic and Enhanced quality; 2 to 5 output variations.
- Best for
- a 3,000-word request free with no account, on text that is already public.
- Main limitation
- nothing published says what happens to pasted text.
- Free tier
- 3,000 words a request, no cap stated, no sign-up.
- Cheapest plan
- none; free only.
- Privacy page on text
- none visible.
- Register controls
- Normal, Clear & Structured.
Why do lists of humanizers for journalists disagree?
Most rank by detector readings, which newsroom policies do not mention, and skip the privacy page, which they do. This page orders by what the vendor states about text, because a reporter's first risk is a source's name on a third-party server, not a detector score, and because the readings change when detectors update while a privacy clause changes only when the vendor edits it. No detector score should be the only evidence in a misconduct case, and no humanizer pass should be the last set of eyes on a published sentence.
Questions about AI humanizers for journalists
Can journalists use an AI humanizer?
Only within the newsroom's policy: AP's standards, read 9 November 2026, treat AI output as unvetted source material and bar staff from using ChatGPT to create publishable content, and the Guardian requires a senior editor's permission and openness with readers for significant AI-generated elements.
Which AI humanizer is safest for unpublished reporting?
None is safe for a source's identity or embargoed material: of the 8 tools here, WriteHuman states text trains its models with no opt-out, GPTHuman and Grammarly state training with a control, and the other 5 say nothing about training, all read 9 November 2026; strip identifying detail before pasting anything.
Does Google penalise AI-assisted journalism?
Not for being AI-assisted: Google's helpful content page, updated 5 October 2026, says using generative AI to produce large amounts of text without oversight is little or no effort and that fabricated bylines are deception, and its February 2023 guidance says quality, not production method, is what counts.
Do journalists have to disclose using a humanizer?
Under most newsroom codes, yes when AI shaped the published text, and under Article 50 of the EU AI Act, text published to inform the public on matters of public interest must disclose AI generation or manipulation unless a person holds editorial responsibility after human review, read 9 November 2026.
What is C2PA and does it apply to text?
C2PA's Content Credentials specification, versions 1.0 to 2.4 as read 9 November 2026, defines signed manifests that record an asset's origin and edits through standard assertions such as Actions and Ingredient; it is designed for images, video, audio and documents, and a humanizer rewrite is an edit a manifest would have to record.
Which humanizer has a journalist mode?
Undetectable AI lists a journalist readability level among its settings, alongside university and high school, read 9 November 2026; no other tool on this page names journalism in its modes, and the setting changes register, not sourcing.
Is a humanizer the same as an editor for news copy?
No: a humanizer rewrites for how text reads to detectors and readers and has no view of what the facts are, while a news editor checks claims against sources, so a humanized paragraph still needs a fact check against the notes it came from.
Does this page recommend a humanizer for newsrooms?
No: it orders 8 tools by what their privacy pages state and lists each one's register controls, cap and price, and leaves the decision to the newsroom's policy; no tool carries a score until the published method's first run.
Sources
- AP, Standards around generative AI, read 9 November 2026: Unvetted source material; no publishable content from ChatGPT; no confidential information into AI tools
- The Guardian, approach to generative AI, read 9 November 2026: Senior editor permission; openness with readers; March 2026 update
- Google Search Central, helpful content, read 9 November 2026: Effort, fabricated bylines as deception, AI disclosures; updated 5 October 2026
- Google Search Central blog, guidance on AI-generated content, February 2023, read 9 November 2026: Appropriate AI use is not against guidelines
- C2PA specification, read 9 November 2026: Manifests, Actions and Ingredient assertions, versions 1.0 to 2.4
- Regulation (EU) 2024/1689, Article 50, read 9 November 2026: Disclosure duty and the editorial-control exemption
- Vendor privacy and pricing pages, read 8 and 9 November 2026: Retention and training wording, plans, caps and modes for the eight tools
Revision log
| Page published. |
Related reading
AI humanizer for legal writing: privacy first, then every defined termAI humanizer for legal writing, ordered by what each vendor's privacy page says about pasted text; 5 of 8.
Best multilingual AI humanizerMultilingual AI humanizer claims compared: the count and the named list each of 13 vendors publishes, read.
AI humanizer for publishers: Google's rules, provenance and editorial controlAI humanizer for publishers: what Google and the EU AI Act say about AI text, what C2PA provenance records.
Best AI humanizer by useBest AI humanizer by use: academic writing, work checked by Turnitin, content writing and more, each page.
Is it safe to use an AI humanizer? What each vendor's privacy page says about your textIs it safe to use AI humanizer tools? What 12 vendors' privacy pages say about keeping your text and training.