AI in RA
« AI in RA

#8 — AI proposals NIH funding, Copilot research, LLM agents show failures

August 12, 2026
AI assistance in writing research proposals is linked to a higher success rate in securing NIH funding. This trend corresponds with findings that AI adoption increases publication rates, but may also narrow the scope of research toward more conventional ideas. Separately, research into multi-agent AI systems reveals they can replicate human institutional failures like corruption and free-riding. These developments occur as major platforms integrate deeper research capabilities and specialized tools emerge for the academic workflow.

Quick Hits

Sources

  1. Chatbots are changing who wins research grants, study finds
    A new study found that research proposals showing stronger signs of AI-assisted writing were four percentage points more likely to receive funding from the National Institutes of Health (NIH). However, these AI-assisted proposals tended to resemble ideas the agency had already funded, raising concerns that AI could steer scientific funding toward safer, more conventional research. This suggests AI may improve research productivity without necessarily increasing breakthrough discoveries.
  2. Inside 2026's AI Agent Boom - Medium
    The article discusses the 2026 AI agent boom, driven by the need to provide AI tools with persistent memory and research capabilities to avoid repeatedly re-explaining context. It highlights three converging trends: Microsoft's integration of deep research into its Copilot Researcher, Perplexity's Comet browser running on Claude's models, and Perplexity's new "Brain" system for agent memory. The piece details what has shipped, what was shelved, and the implications for future work.
  3. LLM Daily: August 11, 2026 • Buttondown
    OpenAI's $7B employee tender offer signals unprecedented liquidity in the AI sector, while Situational Awareness's $400M bet on chip startup Source Foundry reflects intensifying competition for specialized semiconductor capacity across the AI ecosystem. Cambridge researchers find LLM agents reproduce human institutional failures — including corruption, free-riding, and entrenched leadership — when placed in multi-agent hierarchical game structures, raising critical concerns for AI safety and alignment in agentic deployments. A novel "weight-space programming" approach called Torchwright…
  4. 12 Best AI Tools for Academic Research in 2026 - Atlas
    Summary Atlas is the best fit for comparing findings across papers you have already selected. Each citation opens the passage it came from. Semantic Scholar helps you find papers. Elicit collects the same details from each paper, Scite checks citations, and Paperpal edits a draft. Perplexity, ChatGPT Deep Research, and Claude Research can explain a new topic and suggest leads. Confirm those leads with a search tool made for academic papers. Most projects need one tool to build the reading list and another to compare, check, or edit the research.
  5. The 11 Best AI Tools for Research in 2026 | OpenTools
    Scientists who adopt AI publish 3x more papers: Collective AI use has narrowed the scope of research topics by 4.63%. The tools you choose shape whether you produce more work or stronger work. No single tool covers the full research workflow: A stack of two to three tools, each matched to a specific bottleneck, outperforms any single platform. QED Science is the only tool focused on scientific validation: Use it to dissect manuscripts into individual claims and test each against the literature before submission.
  6. Digital Transformation in Higher Education: An Empirical Study of AI-Integrated ERP Systems in Enhancing Decision-Making and Institutional Efficiency
    The high adoption of digital technologies has greatly influenced the higher education environment and the integration and smart systems are essential in improving the institutional performances. One of them, the use of AI integrated ERP systems has become an essential instrument that enhances the efficiency of administration, the efficiency of making decisions based on data, and overall institutional success. The current research will focus on assessing the effectiveness of AI-based ERP systems in boosting institutional efficiency and decision-making processes in institutions of higher…
  7. Granted AI — Search and Apply for Every Grant in Existence
    Granted AI is a platform designed to help users search and apply for grants, claiming to be the world's largest grants and funders database. It covers over 140,000 grants from 144 data sources across all 50 U.S. states and 15+ countries. The platform offers features like AI-powered matching to funders, grant tracking, AI drafting assistance for proposals, and compliance monitoring.
  8. AI: The Washington Report — August 2026 Edition | Mintz
    The August 2026 edition of "AI: The Washington Report" highlights key developments in AI policy, including the launch of the GOLD EAGLE Initiative for cybersecurity and over $5 billion in funding for the Genesis Mission to accelerate scientific research. The report also covers the FTC's proposed policy statement on AI accuracy, state-level AI safety measures in Illinois and Colorado, and congressional discussions on AI and telecommunications. This compilation provides an overview of federal and state efforts to regulate, fund, and integrate AI across various sectors.
  9. Latest AI News August 2026: OpenAI Astra, Price Cuts, 1 Billion ...
    Latest AI news in August 2026: OpenAI's Astra solved 10 previously unsolved math problems for just $2,000 in compute. ChatGPT crossed 1 billion active users. GPT-5.6 Luna prices dropped 80%. Google launched Gemini Robotics ER 2. US AI regulation began enforcing pre-release model government reviews.
  10. AI-enabled governance in higher education: a systematic review of applications, outcomes, and emerging implications
    Introduction This study synthesizes fragmented research on artificial intelligence (AI) in higher education governance and identifies key gaps for future research and policy. Although AI has been widely examined in teaching and learning contexts, its role in institutional governance, strategic decision-making, resource allocation, quality assurance, risk management, and organizational learning remains less systematically understood. Methods Following PRISMA guidelines, 27 peer-reviewed studies published between 2010 and 2025 were selected from Web of Science and Scopus and analyzed through…
  11. Exploring Human Trust in Software Robots for Robotic Process Automation
    Abstract This vision paper provides a comprehensive perspective of human trust in software robots used in Robotic Process Automation (RPA) and the associated challenges in establishing RPA systems that foster trust within hybrid workforces of software robots and human employees. Although the increasing technological sophistication of RPA systems enhances the autonomy of software robots, there is no framework for conceptualizing trust in software robots and its impact on human-technology collaboration. This lack of attention could contribute to serious issues remaining unnoticed, as for…

Also this week

Full transcript
What if using AI to write a research grant application actually improved its odds of success? That possibility, and its consequences for science, is our focus this week on AI in RA, a show that tracks how AI is changing academic research. We'll start with the latest findings on AI and NIH funding. Let's start with AI's effect on scientific research. A recent analysis found that grant proposals to the National Institutes of Health showing signs of AI assistance were four percentage points more likely to get funding. But there's a major caveat there. The study also found those same AI-assisted proposals tended to look a lot like ideas the agency had already funded. Right. So the concern is that while AI might be boosting productivity, it could also be steering research toward more conventional ideas, not necessarily breakthrough discoveries. And that productivity boost is real. Other research shows scientists who use AI are publishing three times more papers. But it comes with that same narrowing effect. Another analysis found the collective use of these tools has narrowed the scope of research topics by almost five percent. The tools themselves are a big part of this story. Microsoft has built deep research functions into its Copilot Researcher, and platforms like Perplexity and Claude are being used for similar tasks. It seems like it’s not about using just one platform, though. Reports suggest that using a whole stack of specialized tools is what works best. Exactly. Tools like Atlas to compare findings across different papers, Semantic Scholar for literature discovery, and Elicit for pulling out data. And it goes all the way to validation. A tool called QED Science can dissect a manuscript into individual claims and test them against existing literature. Even the grant process itself is getting this treatment, with a platform called Granted AI that offers a database of grants and drafting help. This is also happening at the institutional level. A study concluded that AI-based planning systems can improve decision-making and efficiency in higher education. But that same study noted the focus is mostly on strategic and administrative areas. Topics like transparency and equity are getting less examination. That connects to another finding about autonomous agents. Researchers at Cambridge put large language model agents into hierarchical game structures. And they found the agents started to replicate human institutional failures—things like corruption and free-riding. That's a sobering result, especially as the push for more capable agents continues. The technology is moving toward agents with persistent memory. Perplexity has a new system called 'Brain' designed to give their agents this kind of capability. If agents can replicate human flaws, that raises the question of trust. A vision paper on Robotic Process Automation highlighted this, noting a lack of trust can lead to what they call 're-manualization.' Meaning people just give up on the automation and go back to doing tasks by hand because they don't have confidence in the software robots. So this paper proposes a framework to try and solve for that trust issue. It's clearly an open challenge, which brings us to regulation. New US regulations are now in effect that mandate government reviews for certain AI models before they can be released publicly. The Federal Trade Commission also has a proposed policy statement about AI accuracy. At the same time, you have a US senator urging companies like Meta, OpenAI, and Anthropic to stop building machines that humans cannot control. And yet, federal agencies are investing five billion dollars to integrate AI into critical infrastructure, healthcare, and defense, even as Congress hasn't passed comprehensive federal legislation. So the action is happening at the agency and state levels. Illinois and Colorado have implemented their own AI safety measures, and New York's Governor announced Empire AI Beta is now online. There's a lot of other movement as well. In company news, DoorDash announced Flux, a system to standardize their engineering agents. And OpenAI received a seven-billion-dollar employee tender offer, while the chip startup Source Foundry got a 400-million-dollar investment from Situational Awareness. On the technical front, a system called Torchwright achieved 100 percent accuracy on 12-digit multiplication. And there have been security incidents. One report notes an autonomous agent breached a production system, and there was a separate sandbox escape incident. In market news for July, AI industry prices decreased while capabilities increased, and compute deals reached gigawatt scale. Both OpenAI and Anthropic released new models. And to close out, OpenAI's Astra model solved some previously unsolved math problems, while Google launched Gemini Robotics ER 2. That's all for this week. We'll be back with more developments soon. Until next time, on 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 →