#19 — Gemini peer review assistant, federal grant review AI ban

Publishers are now testing automated models to assist human experts in evaluating scientific manuscripts. Where software assistance stops and independent judgment begins is what we examine today on AI in RA, covering artificial intelligence in research administration. Here are the developments. Springer Nature and Google DeepMind are testing a prototype assistant based on Gemini to aid manuscript reviewers, with all files held strictly inside the publisher's technical perimeter. That internal enclosure is the central design constraint. It keeps pre-publication manuscripts from leaking into general model training pools, leaving the actual evaluation choices with human specialists. Though tests indicate a clear division in capability: the models routinely catch verifiable factual errors, but they fail when asked to weigh novelty or scientific relevance. Which connects directly to the friction funding bodies are reporting. Reviewers are submitting machine-generated text in grant evaluations, and agencies are stepping in to halt it. Canada's federal research agencies just clarified that position. Their directive states that institutional software licenses or enterprise vendor contracts do not permit using generative tools on non-public grant proposals. Meaning an enterprise data protection agreement held by a reviewer's university does not waive federal confidentiality rules. The ban on feeding proposals into external models stands regardless of contract tier. In Europe, the framework is expanding as well. The European Commission and the European Research Area Forum issued the fourth edition of their guidelines on generative tools in research, adding rules specifically for autonomous agentic systems. The core of that document remains direct human responsibility, procedural transparency, and explicit records of review assessments across institutions and funding bodies. Yet inside university research offices, the tools are already running daily operations—drafting administrative letters, formatting progress reports, and sorting unit schedules. That cuts turnaround times on routine paperwork, but doing it without standard protocols exposes confidential records. Research administrators are now drafting internal boundaries to keep administrative data compliant with integrity rules. Clinical research teams face a similar tension with patient documentation. Trial sponsors are using automated models to format informed consent text, calculate readability, and tailor phrasing for specific communities. Except these models can generate fluent explanations that distort study protocols or misstate investigational risks. To counter that, sponsors are setting up formal human audits to check every draft against approved study protocols before participants see them. Looking across other developments this cycle, Miro released findings on collaborative artificial intelligence workhubs, alongside studies examining outcomes on corporate productivity and executive choices. Tencent and OpenAI launched models, with OpenAI deploying GPT-6 Cyber to select Daybreak Red customers, while separate teams shipped agent management platforms and tools for scientific work. In government procurement, a Freedom of Information release showed the United Kingdom Cabinet Office uses United States suppliers for ten of its eleven deployed artificial intelligence systems. The Cabinet Office maintains that its internal tools can swap or self-host models when required. Meanwhile, the Prime Minister attended an international artificial intelligence session chaired by President Emmanuel Macron, and vendors introduced a healthcare data connectivity platform focused on administrative and clinical data governance. We will return next week with further updates on institutional policy and research technology. Until next time, on AI in RA.