AI humanizer for scientific writing: what three journal policies require first
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.
The policies come first because they decide whether the pass may be made at all on a confidential manuscript and how it must be declared. Elsevier's policy for journals, updated June 2026 at this reading, exempts basic checks of grammar, spelling and punctuation and requires that "substantive changes to sentence structure or organization of a part of the text" be disclosed in a declaration published with the article (Elsevier, generative AI policies for journals, read 9 November 2026). Springer Nature's editorial policy applies a risk framework in which grammar and language editing is low risk and "suggesting alternative phrasing or framing" requires caution, with the same declaration at both levels, "clearly describe use and confirm author accountability", and states that manuscripts "must not be shared with unsecured or public AI systems" (Springer Nature, AI use in manuscript preparation and Springer Nature, editorial policies on artificial intelligence, both read 9 November 2026). The ICMJE recommendations say journals should require authors to disclose AI-assisted technologies at submission, that writing assistance be described in the acknowledgments, that chatbots cannot be authors, and that authors "should carefully review and edit the result" (ICMJE, defining the role of authors and contributors, read 9 November 2026).
| Policy | What a humanizer pass is under it | Where the declaration goes | Confidentiality rule | Read |
|---|---|---|---|---|
| Elsevier, generative AI policies for journals (updated June 2026) | A substantive change to sentence structure: must be disclosed; grammar and spelling checks are exempt | A declaration statement published with the article | Not stated on the page read | 9 November 2026 |
| Springer Nature, AI use in manuscript preparation | "Suggesting alternative phrasing or framing": caution level; language editing: low risk; both require a clear description of use and author accountability | The manuscript's AI declaration | Manuscripts "must not be shared with unsecured or public AI systems" | 9 November 2026 |
| ICMJE recommendations | AI-assisted writing: disclose at submission; review and edit the output; AI is never an author | Cover letter and the acknowledgments; methods if used for data or figures | Not stated in the section read | 9 November 2026 |
APA and MLA are quoted on the research papers page, and the wording of a declaration by style is on the guide to disclosing AI use.
Eight tools, ordered by what their privacy page states about submitted text, the fact Springer Nature's confidentiality rule turns on: where text is processed or how long it is kept, then stored content the user can delete, then a training clause with an opt-out, then a general clause, then no page. Every figure is the vendor's own, read on the date shown.
What does a rewrite change in a scientific manuscript?
The findings, by small degrees. Scientific prose is hedged on purpose, and a rewrite that reads more naturally tends to remove the hedges, round the numbers and detach citations from their claims.
| What a rewrite can change | Example | What it does to the science | How the method counts it |
|---|---|---|---|
| A causal hedge | "was associated with" becomes "led to" | Correlation reported as cause | Meaning kept, rated 1 to 5 by two editors |
| A scope limit | "in this cohort" disappears | A finding generalised beyond its sample | Dropped or invented facts named in the review |
| A statistic | "p = 0.048" becomes "a significant result" | The reader cannot judge the threshold | Errors introduced per 100 words |
| A unit or dose | "5 mg/kg" becomes "a low dose" | A method cannot be reproduced | Errors introduced per 100 words |
| A defined term | "effect size" becomes "impact" | A technical term becomes an ordinary word | Meaning kept, rated 1 to 5 by two editors |
A declaration names the tool; it does not restore a p-value the tool rounded. The ICMJE's instruction to review and edit the output is the whole job: paste only your own prose, one section at a time, keep figures, units, quotations, citations and the reference list out of the tool, and read every sentence against the original. The guide to humanizing text with citations and quotes sets out the steps.
The published method weights meaning kept at 35 of 100 points and errors introduced at 20, with two editors rating each output blind and naming every dropped or invented claim, and one of its six input genres is a research-style abstract. Those ratings join every block above once a run is published.
Three journal policies read on 9 November 2026, Elsevier's, Springer Nature's and the ICMJE's, each require an author who uses an AI tool on a manuscript's prose to declare it, and Springer Nature adds that manuscripts "must not be shared with unsecured or public AI systems". An AI humanizer for scientific writing is judged first on where the manuscript goes and second on how little of the science the rewrite disturbs, so this page orders eight tools by their privacy statements and ranks none by a score. The overall list is which humanizer keeps your meaning? and the other audiences sit in the best-for hub.
Updated
Short answer
Three journal policies read for this page, Elsevier's on 9 November 2026 and Springer Nature's and the ICMJE's on the same day, each require an author who uses an AI tool on a manuscript's prose to declare it, and Springer Nature adds that manuscripts "must not be shared with unsecured or public AI systems". An AI humanizer for scientific writing is therefore judged first on where the manuscript goes, which is what each vendor's privacy page states, and second on how little of the science the rewrite can disturb. Of the 8 tools on this page, 3 are unsuitable for an unpublished manuscript on their own pages: WriteHuman trains on submissions with no opt-out, Clever AI Humanizer publishes no privacy page, and Undetectable AI's policy dates from 2023 and says nothing about training. No tool here carries a score.
Which AI humanizer for scientific writing states where a manuscript goes?
Price is reported and never ranked. No page read states a flat promise never to train on submitted text, so "with identifiers stripped" is the best case here, not a vendor guarantee. Detector readings under the published method join every block when the first run publishes.
| Order | Tool | Best for | Main limitation | Free tier | Cheapest plan (checked on 8 November 2026 unless dated) | Per-request cap | Named modes | Privacy page on pasted text (read 9 November 2026) | Unpublished manuscript | Review |
|---|---|---|---|---|---|---|---|---|---|---|
| 01 |  Walter Writes AI | Anonymised prose, one section at a time, on the mildest setting | 750 words a request on Starter splits a methods section; the retention clause is written for account data | 300-word trial, no card | Starter, 30,000 words a month, $8 a month billed $96 yearly | 750 (Starter), 1,500 (Pro), 2,000 (Elite, Teams) | Simple, Standard, Enhanced | "processes your text using our own self-hosted AI models"; kept "for as long as necessary to provide you with the Services"; no training statement; last updated 4 September 2025 | With identifiers and data stripped, declared | Review of Walter Writes AI |
| 02 |  StealthGPT | Anonymised prose in 1,000-word pieces under a stated retention limit | No named mode; undated privacy page | None; $1 three-day trial (read 9 November 2026) | Pro, 50,000 words a month, $29.99 monthly (read 9 November 2026) | 1,000 (Pro), 2,000 (Max), 10,000 or more (Enterprise) | None named | Kept no longer than "three (3) months past the termination of the user's account"; no training statement; no date shown | With identifiers and data stripped, declared | Review of StealthGPT |
| 03 |  TwainGPT, now Verva | Prose the author will delete from account history afterwards | 8,000 words a month on Basic is one paper's prose; no free tier, cap or mode stated | None stated | Basic, 8,000 humanizer words a month, $8 a month billed annually (read 9 November 2026) | Not stated | Not stated | "User content may be stored within a user's account history", where users can review and delete it; no training statement; last revised 14 October 2026 | With identifiers and data stripped, declared, then deleted | Review of TwainGPT |
| 04 |  Grammarly AI Humanizer | Drafts already in Grammarly with the training control switched off | Collected information trains models unless the account setting is switched; allowances are AI prompts, not words | Free account, 100 AI prompts a month (read 9 November 2026) | Pro, $12 USD a member, 2,000 AI prompts a member a month (read 9 November 2026) | Not stated | 4 preset voices plus a custom voice from a sample | "We also use information we collect to train our AI models", with a training control in account settings; Superhuman policy effective 6 July 2026 | Only with the training control off | Review of Grammarly AI Humanizer |
| 05 |  GPTHuman | A 300-word abstract trial, then 1,000-word pieces, with the opt-out set | "We may use Content you provide us to improve our Services" unless the account opt-out is used; no modes named | 300 words of output in total (read 9 November 2026) | Starter, 25,000 words a month, $15 monthly or $8.25 a month billed $99 yearly (read 9 November 2026) | 300 free, 1,000 (Starter), 1,500 (Plus), 2,500 (Unlimited) | None named on the pricing page | Kept "for only as long as we need in order to provide our Services"; may train, opt-out in settings; last updated 14 April 2025 | Only with the opt-out set | Review of GPTHuman |
| 06 |  Undetectable AI | Nothing unpublished | A general retention clause dated 2023, no training statement, no modes | 250 words in total (read 9 November 2026) | 10,000 words a month, $5.00 a month charged $60.00 yearly (read 9 November 2026) | Not stated | None named | Kept "for as long as it is necessary for the purposes set out"; no training statement; last updated 18 March 2023 | Unsuitable | Review of Undetectable AI |
| 07 |  WriteHuman | Nothing unpublished through the web tool | Submitted text trains its models, "There is no setting to opt out of this use", and training copies cannot be removed | 3 requests a month, 250 words each (read 9 November 2026) | Basic, 80 requests a month at 600 words, $12 a month billed $144 yearly or $20 monthly (read 9 November 2026) | 250 free, 600 (Basic), 1,200 (Pro), 3,000 (Ultra) | Basic, Enhanced quality | Kept "until you delete it or your account"; trains on submissions, no opt-out; API text "is not used to train our models"; last updated 21 September 2026 | Unsuitable (web tool) | Review of WriteHuman AI |
| 08 | Clever AI Humanizer | Nothing unpublished | No privacy page visible from the vendor site | Free, 3,000 words a request, no sign-up | None (free only) | 3,000 words | Normal, Clear & Structured | No privacy page visible (read 8 November 2026) | Unsuitable | Review of Clever AI Humanizer |
Which humanizer should a researcher choose?
One line per job. Each name links to its full review.
- Choose Walter Writes AI if the prose is anonymised, sections run under 750 words, and a mild setting matters.
- Choose StealthGPT if a stated retention limit matters more than a mode control and 1,000-word pieces fit.
- Choose Verva, formerly TwainGPT, if you will delete the content from account history after each pass.
- Choose Grammarly AI Humanizer only with the training control switched off, for drafts already in Grammarly.
- Choose GPTHuman only with its opt-out set, for pieces under 1,000 words.
- Choose none of Undetectable AI, WriteHuman's web tool or Clever AI Humanizer for an unpublished manuscript.
- Choose no humanizer at all where the target journal or your institution forbids 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.
Sample 01: Walter Writes AI
Walter Writes AI's privacy policy, last updated 4 September 2025, states where text is processed, read 9 November 2026.
- Best for
- anonymised prose, one section at a time, on the mildest setting.
- Main limitation
- 750 words a request on Starter splits a methods section, and the retention clause is written for account data.
- 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.
- Per-request cap
- 750 on Starter, 1,500 on Pro, 2,000 on Elite and Teams.
- Named modes
- Simple, Standard, Enhanced.
- Unpublished manuscript
- with identifiers and data stripped, declared.
Sample 02: StealthGPT
StealthGPT's privacy page states a retention limit and no date, read 9 November 2026.
- Best for
- anonymised prose in 1,000-word pieces under a stated retention limit.
- Main limitation
- no named mode and an undated privacy page.
- Free tier
- none; $1 for three days, read 9 November 2026.
- Cheapest plan
- Pro, 50,000 words a month, $29.99 monthly, read 9 November 2026.
- Per-request cap
- 1,000 on Pro, 2,000 on Max, 10,000 or more on Enterprise.
- Named modes
- none named on the pricing page.
- Unpublished manuscript
- with identifiers and data stripped, declared.
Sample 03: TwainGPT, now Verva
TwainGPT trades as Verva, and its privacy policy, last revised 14 October 2026, states that content may be stored in account history, read 9 November 2026.
- Best for
- prose the author will delete from account history afterwards.
- Main limitation
- 8,000 words a month on Basic, and no free tier, request cap or mode stated.
- Free tier
- none stated, checked on 8 November 2026.
- Cheapest plan
- Basic, 8,000 humanizer words a month, $8 a month billed annually, read 9 November 2026.
- Per-request cap
- not stated.
- Named modes
- not stated on the pricing page.
- Unpublished manuscript
- with identifiers and data stripped, declared, then deleted from history.
Sample 04: Grammarly AI Humanizer
Grammarly's privacy link resolves to the Superhuman policy, effective 6 July 2026, read 9 November 2026.
- Best for
- drafts already in Grammarly with the training control switched off.
- Main limitation
- training is on unless the setting is switched, and allowances are AI prompts rather than words.
- Free tier
- free account, 100 AI prompts a month, read 9 November 2026.
- Cheapest plan
- Pro, $12 USD a member, 2,000 AI prompts a member a month, read 9 November 2026.
- Per-request cap
- not stated.
- Named modes
- 4 preset voices plus a custom voice from a sample.
- Unpublished manuscript
- only with the training control off.
Sample 05: GPTHuman
GPTHuman's privacy policy, last updated 14 April 2025, states an opt-out in account settings, read 9 November 2026.
- Best for
- a 300-word abstract trial, then 1,000-word pieces, with the opt-out set.
- Main limitation
- content may train models unless the opt-out is used, and no modes are named.
- Free tier
- 300 words of output in total, read 9 November 2026.
- Cheapest plan
- Starter, 25,000 words a month, $15 monthly or $8.25 a month billed $99 yearly, read 9 November 2026.
- Per-request cap
- 300 free, 1,000 on Starter, 1,500 on Plus, 2,500 on Unlimited.
- Named modes
- none named on the pricing page.
- Unpublished manuscript
- only with the opt-out set.
Sample 06: Undetectable AI
Undetectable AI's privacy page was last updated on 18 March 2023, read 9 November 2026.
- Best for
- nothing unpublished.
- Main limitation
- a general retention clause dated 2023, no training statement and no modes.
- Free tier
- 250 words in total, read 9 November 2026.
- Cheapest plan
- 10,000 words a month, $5.00 a month charged $60.00 yearly, read 9 November 2026.
- Per-request cap
- not stated.
- Named modes
- none named.
- Unpublished manuscript
- unsuitable.
Sample 07: WriteHuman
WriteHuman's privacy notice, last updated 21 September 2026, is the most explicit on this page, read 9 November 2026.
- Best for
- nothing unpublished through the web tool.
- Main limitation
- submitted text trains its models with no opt-out, and training copies cannot be removed.
- Free tier
- 3 requests a month, 250 words each, read 9 November 2026.
- Cheapest plan
- Basic, 80 requests a month at 600 words, $12 a month billed $144 yearly or $20 monthly, read 9 November 2026.
- Per-request cap
- 250 free, 600 on Basic, 1,200 on Pro, 3,000 on Ultra.
- Named modes
- Basic and Enhanced quality.
- Unpublished manuscript
- unsuitable through the web tool.
Sample 08: Clever AI Humanizer
Clever AI Humanizer's site shows no privacy page, checked on 8 November 2026.
- Best for
- nothing unpublished.
- Main limitation
- nothing states what happens to pasted text.
- Free tier
- free, 3,000 words a request, no monthly cap stated, no sign-up.
- Cheapest plan
- none; free only.
- Per-request cap
- 3,000 words.
- Named modes
- Normal, Clear & Structured.
- Unpublished manuscript
- unsuitable.
Why do lists of humanizers for researchers disagree?
Most rank by a detector reading on one day and never read a journal policy or a privacy page. This page orders by what each vendor states about where a manuscript goes, quotes three journal policies with their dates, and leaves a cell as "no statement" where the vendor does; the corrections log records each change when a policy moves. When the first run publishes, each tool's research-style abstract outputs and their readings on five named detectors will be downloadable with the run.
Questions about AI humanizers for medical and scientific writing
What is the best AI humanizer for scientific writing?
None is usable on an unpublished manuscript as pasted; the three with the most specific privacy statements are Walter Writes AI, which states it processes text on its own self-hosted models, StealthGPT, which states retention of no more than three months past account termination, and Verva, which states content is stored in account history where the user can delete it, all read 9 November 2026, and every pass must be declared under the journal policies quoted on this page.
Do I have to declare a humanizer pass in a journal submission?
Yes under all three policies read for this page: Elsevier requires a declaration for substantive changes to sentence structure, Springer Nature requires authors to clearly describe AI use and confirm accountability at every risk level, and the ICMJE says journals should require disclosure at submission with writing assistance described in the acknowledgments, all read 9 November 2026.
Can I put a manuscript into a public AI tool under Springer Nature's policy?
Springer Nature's editorial policy states that manuscripts, peer review reports and sensitive data "must not be shared with unsecured or public AI systems", read 9 November 2026, and most humanizers on this page are public web tools whose privacy pages state only a general retention clause.
Which humanizers are unsuitable for a manuscript and why?
WriteHuman, whose privacy notice states that submitted text trains its models with no opt-out and that training copies cannot be removed; Clever AI Humanizer, which shows no privacy page; and Undetectable AI, whose policy was last updated 18 March 2023 and says nothing about training, all read in November 2026.
Will a humanizer change a finding or a statistic?
It can: a rewrite that turns "was associated with" into "caused", drops "in this cohort" or rounds "p = 0.048" has changed the science while leaving the sentence grammatical, and no vendor page read for this site states that figures, units or citations are left untouched; the method rates meaning kept by two blind editors.
Is using a humanizer on a paper the same as using an LLM to polish language?
Not under Springer Nature's framework: grammar and language editing is listed as low risk, and "suggesting alternative phrasing or framing" as requiring caution, read 9 November 2026, and a humanizer's whole purpose is alternative phrasing; both levels require the author to describe the use and confirm accountability.
Where do I write the declaration?
Elsevier places it in a declaration statement published with the article; the ICMJE says writing assistance goes in the acknowledgments and the cover letter, and use in data collection or analysis in the methods; Springer Nature asks for a clear description of the use, all read 9 November 2026; the disclosure guide on this site has the wording by style.
Why do lists of humanizers for researchers disagree?
Most rank by a detector reading on one day and never read a journal policy or a privacy page; this page orders by what the vendor states about where a manuscript goes, which is the fact Springer Nature's confidentiality rule turns on, and dates every quotation.
Sources
- Elsevier, generative AI policies for journals, read 9 November 2026: Declaration for substantive changes to sentence structure; grammar checks exempt; June 2026 update
- Springer Nature, AI use in manuscript preparation, read 9 November 2026: Risk framework; language editing low risk; alternative phrasing requires caution; describe use and confirm accountability
- Springer Nature, editorial policies on artificial intelligence, read 9 November 2026: Manuscripts must not be shared with unsecured or public AI systems; human accountability
- ICMJE, defining the role of authors and contributors, read 9 November 2026: Disclose at submission; writing assistance in acknowledgments; AI not an author; review and edit output
- Vendor privacy and pricing pages, read 8 and 9 November 2026: Statements and plans for all eight tools
Revision log
| Page published. |
Related reading
AI humanizer for non-native English speakersAI humanizer for non native english speakers: why detectors flag ESL writing, what Liang 2023 measured, and.
AI humanizer apps for iPhone and AndroidAI humanizer app listings for seven tools, read 9 November 2026: devices, in-app prices against web prices.
Best AI humanizer: 12 tools, one published methodBest AI humanizer by one published protocol: meaning kept, errors added, detector readings and plan limits.
Best AI humanizer by useBest AI humanizer by use: academic writing, work checked by Turnitin, content writing and more, each page.
AI humanizer for research papersAI humanizer for research papers, ordered by the meaning and term controls each vendor states, with journal.