For publishers & institutions
Screen every manuscript for AI transparency.
The question an editor faces is not “was AI used” — no tool can answer that, and the ones claiming to are wrong often enough to ruin careers. It is what evidence exists, whether it agrees with what the authors declared, and what the journal did about it. That last part is the one you get asked about later.
The workflow
Four steps. Screening needs no account and no upload; keeping the record needs an account, because a record has to belong to someone.
Drop the submission package in
The manuscript, its figures, the supplementary files — as a zip or as files. Nothing uploads: the package is opened and read on the machine it is already sitting on, which is the only version of this a journal handling an embargoed paper can use.
Screen a submissionRead the declaration against the evidence
The screen finds the AI-use statement under its own heading, reads whether it declares use or non-use, and compares it with what the files actually carry. It confirms agreement as plainly as it flags a contradiction, and it never concludes that AI was used.
See the outputFollow anything that looks odd
Open a single file for its full signal matrix and edit history, or put two revisions side by side to see precisely what changed between them and what the likely cause was.
Compare revisionsKeep something defensible
Attach the screen to your manuscript reference and it becomes a dated, append-only record: what was checked, under which version of the rules, what the editor decided, and who holds editorial responsibility. Every entry carries the hash of the one before it, so the history is tamper-evident.
Manuscript recordsQuestions it can answer
Each of these resolves to observable evidence rather than an inference.
- Was this figure made in an image editor?
- Software and creator-tool metadata name the application and version where they survive, and a C2PA manifest records the actual sequence of edits.
- Has this image changed since revision 1?
- The hash answers it definitively. If it changed, the comparison shows which signals were lost and what kind of processing most likely caused it.
- Did the authors disclose AI use?
- The screen reads the statement under its own heading and quotes it verbatim, along with the tools and purposes it names. Where no statement is found, it says exactly that — not that AI was used.
- Does the disclosure match what the files carry?
- A manuscript declaring no generative AI alongside figures carrying AI provenance is the one finding worth an editor's time, and the screen states the innocent explanations beside it.
- Are any of these files the same image twice?
- Exact duplicates are flagged across the whole submission by comparing hashes, which catches recycled and re-submitted figures.
- Is this credential genuine and unbroken?
- Signature validation is cryptographic. A valid result means the bytes are unchanged since signing; a failed one means they are not.
- Is our own pipeline destroying provenance?
- The pipeline audit takes the same asset at each stage of your workflow and names the step where credentials stop surviving. Most systems resize on upload, and resizing destroys everything attached upstream.
What this is not
If your primary concern is figure manipulation — spliced blots, duplicated panels, adjusted gels — this is not the tool for it, and we would rather say so now than after you have trusted it.
Not image forensics
We do not analyse pixels for splicing, cloning, or duplicated regions within an image. That is a separate discipline served by dedicated products, and it works on the image content rather than its provenance record. Exact-duplicate detection here compares hashes, which catches identical files and nothing subtler.
Not an AI detector
Nothing here estimates whether an image or a manuscript was AI-generated. Where a producer has declared AI involvement in a signed manifest we report that declaration; where no declaration exists, we say so rather than guessing. No percentage is ever produced.
Not proof a scene is real
A valid credential attests to a file's record — who signed it, and that the bytes are unchanged since. A carefully staged photograph signed by a real camera validates perfectly. Judging what an image depicts remains editorial work.
Not a reason to reject a submission
Most files carry no credentials, because ordinary handling destroys them. Treating absence as suspicion would flag almost every honest author you have. Our measurements show a valid credential surviving one of five routine operations.
Why now
Article 50 of the EU AI Act became enforceable on 2 August 2026, and the machine-readable marking requirement follows on 2 December for systems already on the market. Providers have begun marking output, which means submissions arriving now increasingly carry declarations worth reading — and increasingly arrive stripped of them by the platforms in between.
The practical consequence for an editorial process is not that you can now detect AI. It is that provenance has become a thing you can be asked about, and “we looked” is a much weaker answer than a hashed report with a date on it.
There is a good deal of tooling being described in this space at the moment, some of it promising things we do not believe can be built. We have written up what a verification tool can actually observe — what is readable today, what we would like to add, and where the limits sit — as background reading rather than as compliance guidance.
Everything described here is free and runs in your browser. Saved history, shared cases, vendor records across repeat contributors, and an API are the parts that need servers, and those are what paid plans will cover.