Springer Nature's Review AI Keeps Manuscripts In-House and Judgment Human── The review AI removes the confidentiality barrier to using AI on manuscripts; checking its flags, disclosing its use and deciding on the paper remain human work

On 26 October 2026, Springer Nature and Google DeepMind announced joint research on an AI prototype that supports peer reviewers. When an AI assistant joins peer review, who keeps responsibility for judging the paper? Reviewers and editors do. What the prototype changes is where the manuscript is processed, which is a question of confidentiality, not a change in who decides.

01The prototype moves the manuscript's location, not the people who judge it

Before asking where responsibility goes, it helps to pin down what the announcement actually changes. The striking part is less what the AI does than where it runs. The prototype runs only inside Springer Nature's own Gemini Enterprise environment, Google's AI workspace for companies. Manuscripts do not leave that closed setting.

The people who judge stay the same. Springer Nature's editorial policy states that human accountability cannot be transferred to AI systems. The same policy says AI must not replace scholarly judgement or editorial decision-making. Evaluating a paper and deciding whether to accept it remain the work of reviewers and editors, as before. The announcement itself says the partners are testing what AI can add while being clear about where responsibility remains.

Anyone who uses AI to check documents can take two practical steps from this. Go back to the source and confirm every passage the AI flags. Say in the report that AI was used. Neither step can be skipped, whatever environment the AI runs in.

02The announcement describes a prototype that reads text and figures and flags passages for a closer look

If the people deciding do not change, what does the AI actually do? It is worth defining this strictly within the words of the announcement. The prototype combines Gemini models with the Agent Development Kit, a toolkit Google publishes for assembling AI agents. The AI analyses the text and figures of a manuscript and points to passages that a reviewer should examine more closely.

The announcement names two earlier Google research efforts, ScholarPeer and Deep Review. The new prototype is an attempt to test that line of research in a publisher's peer review.

The boundaries matter. As far as the announcement says, this is not a tool that writes review reports. Nor is it a tool that recommends acceptance or rejection. The AI's job ends once it has marked the passages.

The prototype is still being tested. No date has been given for wider use, and no list of journals. The announcement offers no figure for how often the flags are right.

Why AI was used at all, and why it was built to run only inside the company, can be read as a sequence. Submissions grew, reviewers became scarce, and the NIH and ICMJE confidentiality rules kept public AI tools out.

03Grant review and medical journals restricted AI in review on confidentiality grounds

The prototype keeps manuscripts inside the company. That design has precedents. When generative AI first reached peer review, NIH and ICMJE treated it above all as a confidentiality problem and wrote rules accordingly.

In June 2023 the U.S. National Institutes of Health banned generative AI from its review of grant applications. Its stated reason was that there is no assurance about where text entered into an AI tool goes, or how it is stored and used afterwards.

The International Committee of Medical Journal Editors, an international body of medical journal editors, issues recommendations on submission and review. ICMJE forbids reviewers from uploading manuscripts to AI tools that cannot guarantee confidentiality.

Placed side by side, NIH and ICMJE share a common concern. Both give as their reason where the manuscript or application ends up. I read Springer Nature's decision to run the prototype only in-house as an answer to that concern, given up front.

04Invitations grew 9.8% a year while the acceptance rate fell from 0.54 to 0.44

NIH and ICMJE both kept AI at a distance for confidentiality reasons. Why, then, do publishers still want it? The answer lies in the number of reviewers.

Publons, a service that records peer review activity, published a report on the state of peer review in 2018. It found that review invitations were growing at 9.8% a year. Meanwhile the share of invitations that reviewers accepted fell from about 0.54 in 2013 to 0.44 in 2017. More requests were going out, and a smaller share of them were being taken up.

These figures cover 2013 to 2017, more than eight years before the new announcement. The sources I read give no current acceptance rate. For the present, the evidence is Google DeepMind's own statement: it cites the burden that rising submissions place on reviewers as a reason for the work.

A publisher trying to make up for scarce reviewers with AI has at least two options; here I compare two. It can paste manuscripts into a public AI tool, or it can build a closed environment of its own. The difference between the two is where the manuscript goes. In both cases the reviewer is the one who judges.

The in-house prototype changes the manuscript's destination, which is exactly what the rules objected to. It leaves the person who judges untouched.

05At four AI conferences, 6.5–16.9% of review text may have been heavily edited by LLMs

Alongside the shortage of reviewers, AI has already entered the space around peer review. The extent can be put into numbers.

In 2024, Weixin Liang and colleagues at Stanford examined reviews submitted to four AI conferences, including ICLR 2024. They estimated that between 6.5% and 16.9% of the review text could have been substantially modified by large language models, the class of AI that generates text. That range is an estimate derived from word patterns across the whole body of reviews. It does not show that any particular review was written by AI.

AI is also inside the publisher's own operations. In March 2026 Springer Nature reported that in the previous year nearly 60 AI tools had helped with the handling of more than 1.5 million papers. Its peer review platform, Snapp, is used by more than half of its journals.

While NIH and ICMJE restricted AI in review, LLM edits may already have been present in review text at four AI conferences, and AI was in the publisher's workflows. I read the prototype as an attempt to bring that use into a form in which manuscripts never leave the building.

06A closed environment solves confidentiality only; checking flags and disclosing use stay with the reviewer

Given how far AI has already spread, what does a closed prototype settle, and what stays with people? It helps to separate three issues.

Confidentiality can be solved by where the processing happens

The prototype runs only inside Springer Nature's Gemini Enterprise environment. Within it, the main reason for forbidding AI use, that manuscripts cannot be sent to public tools, disappears. That holds only inside the prototype, though. A reviewer who pastes a manuscript into a public chatbot recreates the original problem.

Whether a flag is right is for the reviewer to check against the manuscript

ICMJE asks reviewers who use AI to make sure its output is appropriate and valid. It also warns that AI output can contain errors, omissions and bias. Part of the reviewer's work therefore becomes comparing each flag with the text. Skip that comparison, and the AI's mistakes flow straight into the review's conclusions.

Using AI is something the reviewer must disclose

ICMJE asks reviewers to tell the journal when they have used AI. Springer Nature's policy says human accountability cannot be handed to AI. Reviewers need a routine step for recording AI use in their reports.

Of the three, the prototype settles only the first. The other two stay with the reviewer whatever environment the AI runs in.

07Whether the flags reduce oversights or draw attention away from unflagged passages has not been reported

If checking the flags remains the reviewer's job, the next question is how flags change the way reviewers read. No test of that has been published for this prototype.

The announcement contains no evaluation of the prototype. Three things are missing.

What follows is my own conjecture. If reviewers read flagged passages with extra care, oversights may fall. If they skim the passages without flags, they will miss whatever the AI missed. Which happens, or whether both happen at once, cannot be known until an evaluation is published. It is a hypothesis, not a fact any source confirms.

Liang's estimate suggests that some review text at four AI conferences already carried LLM edits. Where a flagged manuscript directs a reviewer's attention is, in my view, the next thing publishers need to measure.

Even with an AI assistant in peer review, responsibility for judging a paper stays with reviewers and editors. What the prototype removes is the confidentiality barrier that kept manuscripts away from AI.

Checking the flags, saying that AI was used and deciding on the paper remain in human hands. So that I do not miss what the AI missed, the next time I receive a document marked up by AI I plan to start with the passages that carry no flag.