Peer review is essential to research, but it is also under growing pressure. Researchers face increasing publication volumes, more complex methods and data, and significant demands on the time of qualified reviewers. Against this backdrop, AI is beginning to find a potentially valuable role: not replacing peer reviewers but helping them focus their expertise where it matters most.
On October 26, 2026, Springer Nature and Google DeepMind announced a research partnership to develop an AI peer review assistant. A prototype is currently being tested to help reviewers navigate manuscripts, analyze text, charts, and figures, and identify areas that may deserve closer attention. Specialist scientific judgment, according to the announcement, remains with the human reviewer.
The announcement is notable, but it is also quite preliminary. There is no launch date, list of participating journals, or published information on the prototype’s performance. What it does provide is another signal of where AI-assisted research may be heading. It also resonates with work we have already been exploring at UNU.
From AI-generated reviews to AI-assisted reviewing
Much of the discussion around generative AI and academic publishing has focused on whether researchers should use AI to write peer reviews. A more interesting question may be how AI can help an expert conduct a better and more efficient review.
Springer Nature is not new to AI in publishing. It has already been using AI-supported tools for manuscript screening, editorial evaluation, reviewer recommendations, research integrity checks, and other parts of the publishing workflow. More than 1.5 million papers were supported by its AI-assisted processes in 2025, according to reporting on the publisher’s AI strategy.
The new initiative potentially takes AI one step closer to the intellectual work of peer review itself.
The prototype builds partly on Google’s work with agentic research tools such as ScholarPeer. Rather than treating paper review as a single text-generation task, ScholarPeer uses multiple components to establish context, search the literature, investigate missing comparisons, and examine technical claims before generating a review. Google describes the system as an assistant intended to augment human reviewers rather than replace human judgment.
This is very much the direction we have been experimenting with at UNU.
Revisiting UNU’s Agentic Paper Review System
Earlier this year, we introduced our own Agentic Paper Review System. We first described the approach in January and released the system as open source in June. It uses multiple specialized AI agents to examine papers against defined evaluation criteria, synthesize findings, check claims against the literature, and identify disagreements between different AI models. The system supports the assessment process, but final decisions remain with human reviewers.
The system includes Specialist, Editor, Judge, Librarian, and Fact-Checker agents. It can also run the same papers through different LLMs and surface disagreements between them, providing another signal for where human scrutiny may be particularly important.
Read more about UNU’s open-source Agentic Paper Review System
Although the UNU and Springer Nature initiatives are different, both point toward an important shift. The interesting problem is no longer simply whether an LLM can generate something resembling a peer-review report. It is how AI can be incorporated into a structured review workflow while preserving human oversight.
Assistance can also influence judgment
There is an important tension here.
If an AI assistant highlights particular weaknesses, inconsistencies, or sections requiring scrutiny, it can help reviewers focus limited time on potentially important issues. But the assistance itself can also influence where reviewers direct their attention.
That makes the design of these systems consequential. Researchers need to consider not only whether an AI identifies important problems, but also what happens when it fails to flag something. Reviewers could potentially give less attention to sections the system treats as unremarkable.
Similarly, if many reviewers rely on the same model or analytical approach, the technology could introduce greater consistency while also reducing diversity in how papers are examined. Google’s own ScholarPeer research recognizes review diversity as an important issue and reports reducing the gap between AI-generated feedback and human-level diversity.
These are good reasons to treat AI signals as inputs to judgment rather than substitutes for it.
How generalizable are the results?
There is another reason for caution.
ScholarPeer was evaluated on approximately 1,800 ICLR submissions covering 2020 through 2025. That provides a substantial test bed, but one centered on machine-learning research.
The requirements of reviewing a machine-learning conference paper can be quite different from reviewing work in public health, development economics, environmental science, history, or other disciplines.
Google itself describes ScholarPeer as an experimental research prototype, not a production-ready tool, and says its reviews should not serve as the definitive basis for editorial or publication decisions.
For a multidisciplinary publisher, evaluating performance across disciplines, languages, methodologies, and publication types will therefore be important.
Trust matters as much as capability
Peer review also demonstrates why deploying AI in research cannot simply be a question of model performance.
Unpublished manuscripts may contain novel findings, sensitive information, or intellectual property. Springer Nature says its prototype operates solely within its private, secure Gemini Enterprise environment, and that content used to develop and test the assistant remains within that environment to maintain author and reviewer confidentiality.
That is an important safeguard, but technical isolation is only one part of the governance question.
Research organizations adopting similar systems will need to consider where information is processed, what the AI is permitted to access, whether external information can be retrieved, how outputs are stored, how AI involvement is disclosed, and how errors or disagreements are handled. Most importantly, they will need to remain clear about who is accountable for the resulting judgment.
A broader opportunity for research
Peer review is only one application of this design pattern: an AI assistant that examines material against defined criteria and surfaces issues for a human to judge. The same approach could support literature analysis, policy analysis, evidence synthesis, grant assessment, and other forms of research quality assurance.
This is why UNU’s Agentic Paper Review System was designed to be domain-agnostic. Its criteria and scoring can be configured for academic papers, grant proposals, project reports and policy briefs. The Springer Nature and Google DeepMind announcement gives us a timely reason to revisit an idea we have already been exploring. It describes a prototype still in testing, so the practical questions are open:
- Reviewer guidance: How will reviewers be told when and how to use AI assistance, and which disciplines and journals will use it first?
- Author transparency: Will authors know that AI was involved in assessing their manuscript, and can they opt out?
- Measurement: How will errors, missed issues and false flags be measured, and by whom? Has the tool been evaluated independently and across disciplines?
- Reviewer agency: Must reviewers engage with the AI’s flags, or can they set them aside? Are overrides recorded?
- Editorial visibility: Will editors see the AI output, or only the reviewer’s final assessment?
- Criteria and bias: Who sets the criteria and scoring, and does the tool systematically favor certain styles, topics or institutions?
- Depth of thinking: Does the tool help reviewers think more deeply, or simply narrow where they look?
The question is shifting from “Can AI review a research paper?” to “How do we design research workflows in which AI helps people examine evidence more effectively, while humans remain responsible for expertise, context and judgment?”
The answer will depend on more than better models. It will take careful workflow design, transparency, evaluation and governance.