Can You Upload a Manuscript to AI? Journal Policies, Explained

If you are a peer reviewer or an editor, you have probably asked a version of this question recently: can I paste a manuscript into ChatGPT? In most cases the answer is no, at least not the manuscript itself. Most major publishers now have AI policies that treat a submitted manuscript as confidential, and a general-purpose cloud chatbot does not meet that standard.

However, the rules differ from one publisher to the next, and the wording keeps changing, so a plain "no" is rarely the whole answer. The sections below explain what these policies are protecting and how to get AI help on a review without crossing the line.

What the policies are protecting

Peer review depends on confidentiality. When an author submits a manuscript, they hand over unpublished work to a small number of trusted people, with the expectation that it will not be copied, shared, or entered into other systems.

Publishers inherited that expectation mainly from two sources. The ICMJE recommendations describe peer review as confidential and keep the manuscript under embargo until publication, while COPE guidance treats confidentiality as a right of the author and an obligation of the reviewer.

Generative AI sits awkwardly against both, for three reasons:

You do not control the data once it is uploaded. Consumer chatbots may retain inputs, use them for training, or expose them to human reviewers, depending on settings you do not fully control.
The manuscript is not yours to share. Even if you are comfortable, the author never agreed to it.
It can distort the record. If AI text ends up in a review or a decision, the journal may not be able to say who wrote it.

The common policy across publishers is therefore that reviewers may not upload a manuscript, or any part of it, into a tool that does not guarantee confidentiality. In practice this rules out consumer cloud chatbots.

What the major publishers say, in general terms

Policies differ in tone and detail, but most converge on a few points. I will not quote a specific clause, because any clause I quote may already be outdated. Treat the following as recurring themes, and always read the policy of the exact journal you are reviewing for. Most start from confidentiality: reviewers are reminded that the manuscript is privileged and that uploading it to an external service can breach that duty. Many require disclosure, so a reviewer may have to declare the use of any AI-assisted tool even when no text from the manuscript was shared. The objections are usually aimed at the tool rather than the task, which means they rarely cover spell-checkers or AI that runs locally on your own machine. A few journals allow AI use only where the author has agreed, or where the tool runs under the publisher's own contract.

The safe default that follows from all of this is that the manuscript, and anything taken from it, should not leave your machine.

Where the line falls

It helps to separate three things that people lump together under the phrase using AI:

Activity	Usually fine?	Why
Asking a general AI a question without the manuscript ("what does a two-tailed Mann-Whitney test assume?")	Generally yes	No confidential content is shared
Pasting text, figures, or data from the manuscript into a cloud tool	Usually no	Breaches confidentiality as most policies define it
Running an AI model locally, with the manuscript never leaving your device	Often allowed, but check your journal	The confidentiality concern is what the policy is really targeting

That third row is the part that changed recently. A local model and a cloud chatbot are often spoken about as if they were the same thing, but only the second one sends your manuscript off the machine.

Why local changes the analysis

A model that runs on your own laptop, reading a PDF from your disk and writing its output back to your disk, never transmits the manuscript anywhere. There is no service and no account, so there is nothing to retain. For a reviewer, that maps closely onto the confidentiality the policies are protecting.

Two points are worth keeping in mind. Local does not automatically mean permitted, because a journal's policy may still require you to disclose that you used AI assistance at all; disclose whenever you are unsure. Local models are also not a guarantee of quality, because a small model can be confidently wrong, and the reviewer, not the model, owns the review.

This is the gap Peerscy was built to fill. It is a desktop app that runs a local LLM in process, offline, so the manuscript never leaves your machine while you still get real review work done.

How to review a paper with AI without uploading it

A workflow that stays within typical policies:

Keep the manuscript local. Attach the PDF to a tool that runs offline. If you can switch off Wi-Fi and it still works, confidentiality is not in question.
Read and structure locally. Ask for a summary, a methods audit, or a list of claims to verify, all of it against the local file.
Verify the deterministic things. Reported statistics and reference lists can be checked without any cloud service: recompute the p-values from the test statistics, and resolve the reference DOIs to catch retractions.
Disclose. If the journal asks whether AI tools were used, say so. "Offline assistant for consistency checks, all judgments mine" is a clean and honest answer.
Own the verdict. Treat what the model produces as a draft to check. The conclusion is still yours to write.

On the deterministic side, Peerscy recomputes reported statistics (t, F, chi-squared, r, and z, with their degrees of freedom and p-values) and flags inconsistencies, and it resolves the reference DOIs in a manuscript to flag retractions, all of it offline. You can read more about the privacy model, or, if you build with agents, the local MCP server exposes the same tools with nothing hosted.

A checklist before you upload anything

Does the journal's AI policy mention confidentiality or the manuscript specifically?
Is the tool confidential, or could my input be retained, reviewed, or used for training?
Am I uploading the manuscript, a figure, a table, or data, or just asking a general question?
If I use AI at all, does the journal require disclosure?
Could this task run on my own machine instead?

The bottom line

These policies are really about confidentiality. Consumer cloud chatbots fail that test, which is why the answer to whether you can upload a manuscript is usually no. A model that runs on your own machine never puts the manuscript at risk to begin with, and that is why local, offline tools are becoming a practical way for reviewers to stay productive and compliant.

Whichever tool you use, read the policy for your journal, and disclose when you are unsure.