AI in RA
« AI in RA

#10 — NIH AI grants less novel, 1% AI authorship .edu, open-source AI parity

August 26, 2026
Research indicates a correlation between AI use in NIH grant applications and a decrease in the novelty of proposed projects. This development in research funding is contrasted by data showing low AI adoption on academic domains compared to commercial ones. Meanwhile, the AI landscape is shifting as open-source models match proprietary performance and costs drop. These changes prompt new governance frameworks in higher education and evolving national safety policies.

Quick Hits

Sources

  1. AI Intelligence Briefing — August 20, 2026 • Buttondown
    This edition of the AI Intelligence Briefing covers several key topics: AI's role in NIH grant success leading to less novel research, a study finding major AI companies lack sufficient safety measures, OpenAI's temporary pause in development for security, and the quick circumvention of Claude's invisible watermarks. It also details Google's new AI study tools in Search and Gemini, and a study on LLM serving workloads. The overarching theme highlights the tension between AI capabilities and containment, underscoring the need for robust institutional policies regarding AI adoption and…
  2. AI Intelligence Briefing — August 22, 2026 • Buttondown
    💡 Signal: This week's signal is a governance inflection point. The White House is simultaneously expanding its AI safety framework to cover open models while its own national security strategy omits open-weight AI from the critical technology list — sending mixed signals to enterprises and universities alike. Meanwhile, Pew's data showing .edu domains at just 1% AI authorship (vs. 10% for .com) offers a narrow window before AI-assisted writing normalizes in academia. The vLLM kernel divergence paper is a quiet bomb for anyone running quantized inference in production: your model outputs may…
  3. Latest AI Developments: August 2026 Update - Local AI Zone
    August 2026 marks a turning point: The pace of model releases has outrun anyone's ability to fully test them (11+ models in 20 days from 5+ providers). Anonymous frontier models achieve production adoption within hours. Open-source models match proprietary performance. The cost per intelligence unit dropped ~50% across multiple tiers. And the agent revolution moved from experimental to essential infrastructure.
  4. AI in education: From experimentation to institutional impact ...
    Microsoft's 2026 AI in Education report indicates that AI use is widespread, with 92% of surveyed students and education leaders and 88% of educators using AI for school-related purposes. The challenge now is to move AI beyond individual productivity to institutional reliability across teaching, learning, research, and operations. Microsoft 365 Education aims to bridge this gap by integrating relevant institutional context and safeguards, helping institutions transition from isolated AI experimentation to practical, governed use.
  5. Responsible AI in Polish Higher-Education Administration under the EU AI Act: High-Risk Use Cases, Fundamental-Rights Impact Assessment, and Institutional Accountability
    Artificial intelligence (AI) is entering university admissions, assessment, proctoring, student support, staff management, and administrative communication. Yet neither the label ‘educational AI’ nor a human signature at the end of a workflow determines the applicable legal regime. This article asks how Polish higher-education institutions (HEIs) should classify and govern administrative AI under Regulation (EU) 2024/1689, as amended by the Digital Omnibus on AI, the General Data Protection Regulation (GDPR), and Polish public and higher-education law. A doctrinal and legal-design analysis…
  6. Beyond AI detection: African universities must choose transparency ...
    Artificial intelligence is here to stay and African universities must adapt accordingly. Rather than relying on imperfect AI detectors, institutions should promote transparency, ethical AI literacy and assessment methods that reward critical thinking and genuine understanding.
  7. Webinar | Grant Writing in the AI Era: Practical Strategies for Better Outcomes
    This webinar, organized by DevelopmentAid in collaboration with Paulius Yamin, will explore how Artificial Intelligence can be used effectively in grant writing to improve proposal quality and speed without compromising compliance or credibility. It aims to provide bid managers, proposal writers, consulting firms, and NGOs with practical strategies and a framework for integrating AI into the grant writing process. The session will cover opportunities and risks, helping participants harness AI capabilities while avoiding common pitfalls.
  8. Singapore agentic AI adoption doubled in 2026, with 10% of ...
    A new ServiceNow study reveals that Singapore's agentic AI adoption doubled in 2026, with 10% of enterprises redesigning workflows, pushing the nation's AI maturity score to 53 out of 100, above the global average. While Singapore leads in most AI maturity dimensions, it lags in AI-enabled workflows, indicating a gap between foundational investments and embedding AI into daily operations for business returns. The study emphasizes that organizations that redesigned work processes around AI were nearly four times more likely to report significant productivity gains.
  9. Workday Introduces AI Research Team Dedicated to Advancing Reliable, Trustworthy, and Efficient Enterprise AI
    Workday, Inc. announced the establishment of Workday AI Research, a technical research team focused on advancing reliable, trustworthy, and efficient AI for enterprise applications. This team's findings will help shape Workday's AI development and contribute to the broader research community. Workday is also introducing the Workday AI Research PhD Fellowship to foster collaboration with academia and support doctoral students working at the intersection of AI and enterprise software.
  10. Development and validation of the trust in AI scale (TAIS)
    In everyday life, users increasingly interact with AI systems. Despite the importance of trust in AI as an influencing factor for this interaction, there is a shortage of validated scales to reliably measure users’ trust. In this paper, we present a theory-driven development and validation of the Trust in AI scale (TAIS) that consists of the subdimensions ability, integrity, transparency, unbiasedness, vigilance, and global trust. To validate the scale, we conducted two studies. In study 1 (n = 883participants), we derived 57 items from theory and existing scales, for which an exploratory…

Also this week

Full transcript
Using AI to write NIH grant proposals appears to correlate with less novel research. On AI in RA, we examine the effects of artificial intelligence on research and academic work. We begin with that story. The entire artificial intelligence field seems to be in a state of rapid acceleration, both in terms of what the models can do and the rush to figure out how to govern them. And that acceleration is fueled by a couple of key changes. Open-source models are now performing at the same level as proprietary ones, and the cost for a given unit of AI intelligence has dropped by about half. Which explains why we're seeing AI agent systems go from being experimental to becoming basic infrastructure components for companies. But that pace creates a problem. We saw a period where five different providers released more than 11 new models in just 20 days. There's no way to do comprehensive testing on that kind of schedule. So you get this situation where even anonymous models, with no clear origin, are being put into production use within hours of their release. This is also changing how research gets done, and not always in the ways you'd predict. What do you mean? An analysis of National Institutes of Health grant applications found something interesting. When researchers used AI tools, their proposals were correlated with being less novel. So the tool intended to expand creativity might actually be narrowing the scope of scientific inquiry. That seems to be the suggestion. But at the same time, you have organizations like DevelopmentAid creating webinars to teach people practical strategies for using AI in grant writing. It’s being adopted even as these questions come up. Though the data shows a split. Pew research indicates AI-assisted content is still very low in academia—only 1% on .edu domains, versus 10% on commercial .com domains. To try and bridge that gap, you see efforts like Microsoft 365 Education planning to build institutional context and safeguards into its AI tools. The idea is to make them more reliable for teaching and research. This all points to the need for more structured governance, which is also becoming a subject of academic work itself. Right. A paper looking at Polish higher-education institutions developed frameworks for how to classify and manage administrative AI under EU law. And they produced some concrete models for doing this, like a use-case classification matrix and something called the HEI Administrative AI Four-Gate Model. It’s about creating auditable ownership. The analysis also touched on themes we’ve discussed before: the need for meaningful human authority over these systems and for their decisions to be understandable and contestable. And under the EU AI Act, this is not just a suggestion. Certain uses in education, like for admissions or student evaluation, could be classified as high-risk systems, which brings a whole set of stricter legal duties. At the government level, there are some mixed signals. The White House is extending its AI safety framework to open models, but it also left open-weight AI off its list of critical national security technologies. Meanwhile, enterprise adoption is growing. One study from ServiceNow found that the use of agentic AI doubled among enterprises in Singapore in 2026. And that adoption pushed the country's AI maturity score above the global average. The organizations that saw the biggest productivity gains were the ones that actually redesigned their workflows around the AI. It's not just about plugging it in. Which also means understanding how people interact with these systems. Researchers have now developed a validated scale for this, called the Trust in AI scale, or TAIS. It measures trust across different dimensions—a system's ability, its integrity, transparency, and its unbiasedness. You're also seeing companies formalize this, like Workday creating a dedicated AI research team for reliable enterprise AI. And the same push is happening elsewhere. A recent report advises that African universities need to adapt their strategies to integrate AI as a permanent part of their academic work. There have been a few other developments recently. Google expanded access to its AI learning tools for university students in Africa. There’s also a finding that some of the largest AI companies don't have sufficient safety measures in place. We saw OpenAI pause a development for security reasons, and the invisible watermarks in Claude's models were circumvented pretty quickly. Another technical finding showed that outputs from large language models can sometimes depend on the specific low-level software kernel that gets loaded, which introduces another layer of variability. On the legal side, there's been more analysis of the EU AI Act. Researchers mapped the boundaries between its requirements and GDPR's, and clarified that using AI for emotion inference in education is generally prohibited under the act. That same analysis also clarified which types of private universities are subject to certain assessments and interpreted the December 2027 delay for some of the act's obligations. We'll track these stories and more in our next episode. Thanks for listening to AI in RA.

From our community

Builders Group

A peer working group for research administrators who are actively building AI into their offices. Some members work alongside IT teams; others are piecing it together on their own. What unites the group is the work of implementation itself:

RSVP →

From our community

Join AI in Research Administration

Join the community →