For publishers, editors, agents and prizes
See which passages of a submission read as machine-written, and why, with error rates published by genre. The model runs in your browser, so the manuscript is never uploaded, stored or used for training.
500 words free · no account · the text never leaves this browser
Paste a chapter, an abstract or a submission, or drop a .docx or .pdf. It never leaves this browser.
Open the checker: 500 words free, nothing uploadedExample report
Illustrative example · invented text
An invented sample, run through Ghost Act exactly as shown. Scored as a whole it leans human. The mixed-authorship check splits it in two: the opening at about 0.03, the last paragraph at about 0.93.
Calibrated ensemble
25% AI likelihood for the whole text. 15 passages marked, each with the reason.
What to take from itA single score would wave this chapter through. Ghost Act shows where the writing changes and why: every flag sits in the final paragraph. That is a passage to raise with the author, not a verdict on the book.
Four hundred manuscripts, and a promise not to upload them
The submissions window closed on Friday. There are 400 manuscripts in the queue and two of you to read them. Your contract template, like the Authors Guild’s new model clauses, asks authors to disclose AI-generated text and promises you will not upload their work to consumer AI systems without permission.
A few sample chapters feel wrong: fluent, even, oddly weightless in the back half. Every detector you have tried wants the file on its servers. You would be checking the promise by breaking it.
With Ghost Act the check runs in your browser. You see that the opening reads like the author’s own voice and the last pages read differently, with a reason beside every flag. So you write to the author about four specific paragraphs and ask about their drafts, instead of sending an accusation. The manuscript never left your laptop.
The problem
Retailers are rejecting AI-generated books in bulk and journals are seeing LLM prose in a growing share of papers. Yet the Authors Guild says no publisher runs a scanner or verifies authors’ AI statements before signing, and journal policies tell editors not to upload manuscripts to AI tools at all.
of self-published submissions rejected by Kobo in 2025; over 80% of those were judged AI-generated or of very poor quality.
The Bookseller (2026)of 2024 biomedical abstracts at least were processed with an LLM, rising to 40% in some subcorpora.
Kobak et al. (2025)machine-written stories reached Clarkesworld before the magazine closed submissions in February 2023.
TechCrunch (2023)AI-written articles at one technology news site needed corrections after publication, some for basic factual errors.
Engadget (2023)How Ghost Act helps
“We promised not to upload it”
The language model runs in your browser. Unpublished work is not uploaded, not stored on a server and never used for training, and saved reports stay in that browser. Screening no longer means handing an author’s book to a third party.
“Assisted, or generated?”
The mixed-authorship check splits a manuscript where the evidence changes. A whole-book score can lean human while one chapter does not, and that chapter is where disclosure clauses and “de minimis” thresholds are decided.
“What do I say to the author?”
Every flag names its cause: LLM vocabulary, stock phrases, transition stacking, rule of three, uniform rhythm. Put particular paragraphs to the author and ask about their drafts, rather than confronting them with a percentage.
“What if we wrongly accuse a real author?”
Tested on genres it never trained on, Ghost Act wrongly flagged 0% of human book prose, 2% of news, 4% of academic abstracts and 14% of encyclopaedic writing. We publish the weak numbers too.
“Four hundred submissions, two readers”
Pro checks 500,000 words a month with batch reports, and an MCP server lets your submission system request checks. Borderline work is marked “Uncertain” and goes to a human reader; nothing is rejected automatically. Checks through the MCP server run on Ghost Act’s server in Western Europe: analysed in memory, never stored or used for training.
“Do we have to label it?”
Plain-English explainers on when AI text published to inform the public must be disclosed under the EU AI Act, why human editorial review is the exception, and why “no watermark found” on its own tells you nothing at all.
Compared
Most AI detectors run in the cloud. That is fine for a published blog post, and awkward for an unpublished novel under contract or a paper under peer review.
| Feature | Ghost Act | Typical cloud detector |
|---|---|---|
| Where the manuscript goes | Stays in your browser: never uploaded, kept on a server or used for training | Uploaded to the vendor; often kept, and training use depends on the terms |
| What you get | Which passages read differently, and the reason for each flag | Usually a document-level percentage, sometimes with highlights |
| Borderline work | An explicit “Uncertain” verdict | A number either way |
| Error rates | False flags published per genre, plus recall on models it never saw | Typically one headline accuracy figure from the vendor’s own tests |
| Watermarks | Open-scheme tests, and when “no watermark” means nothing | Rarely covered; a missing watermark is rarely explained |
| Rewriting | Never rewrites or “humanises” text | Some also sell a humaniser |
Paste a chapter or drop a .docx, .pdf or .txt file, or send a whole batch on Pro. It is read in your browser; 500 words are free.
See which passages differ from the rest and the reason for each flag. “Uncertain” means read it closely, not reject it.
Put specific passages to the author, ask about drafts and process, and file the printable report with your editorial notes.
Every figure on this page comes from Ghost Act’s published validation. How we measure
For editorial teams
£100/month500,000 words a month: roughly five or six full-length novels, or around a hundred journal manuscripts, with batch reports and an MCP server for your submission pipeline.
Questions
Not with certainty. The Authors Guild says it is not aware of any truly reliable test, and every detector makes mistakes. What a detector can do is show which passages read like LLM output and why. Ghost Act catches 81% of AI text at 3% false flags on held-out data, and marks borderline work “Uncertain”. Use it to decide what to read closely and what to ask the author.
Yes. Ghost Act runs its language model in your browser, so the text is not uploaded, not stored on a server and never used for training. Saved reports live only in that browser. You can check a sample chapter or a whole manuscript without handing it to a third party.
Elsevier tells reviewers and editors not to upload a submitted manuscript, or any part of it, into an AI tool, because it may breach the authors’ confidentiality; Springer Nature says similar. Ghost Act does not upload the text: it is analysed on your device. Whether that fits a particular journal’s policy is for the journal to decide, so check with your editorial office. Checks sent through Pro’s MCP server are different: the text goes to Ghost Act’s server, where it is analysed in memory and never stored, so ask about that route too.
KDP requires authors to disclose AI-generated text, images and translations, meaning content an AI produced, even if they substantially edited it afterwards. AI-assisted work, the author’s own writing refined with AI tools, need not be disclosed. Ghost Act’s mixed-authorship check shows which parts of a manuscript read differently from the rest, which is where that line gets drawn.
Article 50(4) requires deployers to disclose AI-generated text published to inform the public on matters of public interest, unless it has undergone human review or editorial control and someone holds editorial responsibility for it. It applies from 2 August 2026. Ghost Act explains the rule; it does not make you compliant. This is general information, not legal advice.
On genres it never trained on, it wrongly flagged 0% of human book prose, 2% of news, 4% of academic abstracts and 14% of encyclopaedic writing. On an independent set of co-written documents it flagged 2% of human-only texts and 8% of mostly-human co-written ones. It is weaker on human text that an LLM has rewritten, catching 33%. Full results are on the Methods page.
It is tested against 24 AI models, including GPT-6, Claude Opus 5.5, Gemini 3.1 and Llama 4, and catches 82% of text from models it never saw in training; on an independent benchmark its AUROC is 0.83. It runs open-scheme watermark tests, but it cannot read Claude’s or Gemini’s keyed watermarks, and it explains when “no watermark found” means nothing.
Yes, on Pro: 500,000 words a month for £100, with batch reports and an MCP server so your submission system or agents can request checks. Text sent through the MCP server is analysed in memory on Ghost Act’s server in Western Europe and never stored or used for training. For occasional use, word packs start at £4.99 for 5,000 words.
Start with 500 words free, no account. See which passages read differently and why, then decide what to ask the author.
Check a manuscript free