#16 — AI tested on COPE ethics, automated trial screening, NSF grant terms

Can language models reliably navigate the ethical disputes that arise in scientific publishing? That is what we examine today on AI in RA, tracking artificial intelligence across research administration. We begin with publication ethics. A recent study tested Gemini 2.5 Flash and DeepSeek-V3.2 on publication ethics disputes, comparing their recommendations against historical decisions from the Committee on Publication Ethics. They ran eleven anonymized COPE cases through both models and graded the outputs against human consensus with a structured rubric. Both models produced recommendations that tracked COPE's historical guidance. In several cases, they even suggested steps beyond the original record, like contacting secondary institutional bodies or third-party collaborators. Yet the evaluation also flagged omissions and contradictory advice across outputs. That inconsistency led the authors to conclude automated models cannot replace human editorial committees in adjudication. That tension between automated throughput and oversight is showing up in workflow metrics elsewhere, specifically in clinical trial recruitment. Ochsner reported concrete figures on their algorithmic screening rollout: screening capacity rose by 41 percent, while manual screening time dropped by roughly 75 percent. At the same time, the volume of patients tracked for future protocol eligibility expanded by a factor of 3.6. Shifting from hospital operations to federal policy, the National Science Foundation confirmed it is updating documentation standards across solicitations, proposal guides, and award paperwork. Pursuant to an executive order, NSF grant materials will replace references to artificial intelligence with the term Super Intelligence. Intellectual property policies around computational tools are also shifting at the National Institutes of Health. Officials from the NIH Office of Technology Transfer are presenting at the Licensing Executives Society meeting in San Diego. Director Tara Kirby is speaking on licensing biomedical data in machine learning workflows and leading a session on access planning. Alongside that, Special Advisor Steven Ferguson will present on Bayh-Dole Act compliance, risk mitigation, and diligence protocols for federally supported discoveries. Several related initiatives were announced this cycle as well. UC Davis Health started a clinical trial assessing HiDO Health's automated medication dispenser. AstraZeneca opened recruitment for a regulatory writing manager focused on machine-assisted documentation, and an institute launched a training program on artificial intelligence use in manufacturing. For research tooling, the AI Institute in Dynamic Systems developed methods to model and simulate complex dynamic systems, while the National Science Foundation outlined a 75 million dollar funding opportunity for Super Intelligence-Native Instruments. That follows the administration's announcement of six billion dollars in cross-sector scientific investments presented at the Science: A New Golden Age summit. On immigration and workforce policy, the Department of Homeland Security proposed fees up to 100,000 dollars for the Optional Practical Training program. Finally, researchers proposed comprehension audits as a mechanism for frontier model development oversight, while another team built a locally deployed software framework for administering university physics experiments and instruction. We will return next week with more updates on research policy and systems. From AI in RA, thanks for listening.