#14 — Automated UKRI grant scoring, ethics triage oversight, Lyn assistant

Algorithmic scoring is entering the review pipeline to evaluate and filter public research proposals. Examining that operational shift is our focus today on AI in RA, covering artificial intelligence across research management. Here are the stories. When research ethics committees bring large language models into protocol triage, the standard assumption has been that keeping a person in the loop acts as a corrective buffer. Yet the research indicates human intervention can actually widen model misalignment. The models rely on observable proxies instead of the actual ethical standards, which creates an immediate proxy-target gap. And once human reviewers begin interacting with those models, secondary issues appear. Reviewers encounter sycophancy, where the model aligns with their inputs, or over-reliance, where the reviewer defers to the automated output. Not to mention hallucination and reward hacking. To push back against that dynamic, the researchers point toward specific interventions: structured prompting, adversarial debate protocols, calibration drills, and retrieval grounding, alongside targeted training for reviewers. That proxy question is also surfacing in grant allocations. A preprint on UK Research and Innovation applications evaluated automated scoring models designed to screen out lower-ranked submissions. Which connects directly to field reports of proposals receiving automated rejections. When funding bodies introduce algorithmic screening, the debate centers on where the decision boundaries are set and how those criteria are implemented in standard peer review. The World Health Organization took up similar governance questions in a report on artificial intelligence in health research. Their framework details explicit requirements for participant consent, methods for detecting algorithmic bias, and rules for equitable benefit sharing. They also examined institutional capacity—specifically whether ethics committees possess the operational resources to evaluate protocols that contain machine learning systems in the first place. Understanding how human review actually functions across disciplines sheds light on why standardized oversight is difficult. An analysis of the Swiss National Science Foundation evaluated 39,280 peer review reports from 2016 through 2023. That covered 1.3 million sentences across more than 11,000 proposals. The supervised model identified distinct patterns: reviews in the humanities and social sciences were systematically longer and contained more critical sentiment than reviews in technical fields and life sciences. Which shows baseline variation in qualitative review before automation is even applied. Meanwhile, for portfolio-level analysis, some funders are opting for simpler architectures. In Thailand, researchers tested an explainable text-classification system to audit an area-based funding portfolio. They used term frequency-inverse document frequency paired with a linear support vector machine across 629 project records categorized by poverty metrics. It reached a weighted F1 score of 0.7798 across five-fold cross-validation without using a black-box model. On the administrative side of research, vendor tooling is moving into operations. Streamlyne launched an assistant named Lyn within its platform to draft proposal budgets, answer policy questions, and redline contracts against local rules. With clear technical constraints: it operates within institutional data perimeters using role-based access, and requires administrative approval checkpoints before any proposal can be submitted or a contract executed. Professional associations are formalizing their boundaries around these tools as well. NCURA and SRAI updated their guidance to allow generative tools for drafting and editing, on the condition that assistance is disclosed and the human author remains accountable for accuracy. Both organizations also instituted outright bans on automated meeting bots recording conference sessions and organizational meetings. Shifting to policy items on the radar, Donald Trump plans to create a political appointee board to oversee National Institutes of Health grant awards, while the White House drafts an executive order establishing an external review panel for NIH proposals. Concurrently, tracking organizations report that administration policy shifts have led to the disruption or termination of thousands of research grants nationwide. In international governance and security, Sam Altman addressed the United Nations to request cross-border regulation and rapid incident reporting mechanisms. In Australia, the Prime Minister stated that OpenAI delayed too long before disclosing a data breach. On the defensive front, Proofpoint introduced agentic security systems for enterprise and collaboration tools. Meanwhile, researchers tested platform defenses by simulating an agentic attack on Hugging Face, as frontier models demonstrated autonomous, multistage cyber operations against standard systems. For data center infrastructure, Nvidia established a qualification program for DSX-compliant facilities, and partnered with Vertiv to propose measuring operational efficiency in tokens per watt. At the policy level, the European Union rolled out an energy label to monitor data center water and electricity consumption, while hardware providers announced backing for 800-volt direct-current architectures. In domain-specific applications, a study mapped the concrete uses and mechanical limits of machine learning within manufacturing and materials research workflows. Another publication introduced a paradoxical leadership framework to balance workforce coordination alongside automated systems, while education data showed an asynchronous professional development course shifted K-12 teacher comfort with these platforms. Finally, researchers published an ethical analysis focused specifically on agentic systems in clinical healthcare, as commercial operations in plant science and agriculture accelerated the deployment of automated sensors and robotics across physical field facilities. We return next week with more updates on policy and institutional governance. From AI in RA, thank you for listening.