New AI Agents Pose “Existential Threat” to Grant Awarding

The rise of more sophisticated artificial intelligence agents poses an “existential threat” to the way research funding is awarded, experts have warned, saying autonomous AI systems could further flood grant competitions with applications, making it harder to identify the best ideas.

Speaking during a webinar organized by the League of European Research Universities, Geraint Rees, vice provost for research, innovation and global engagement at UCL, said a new wave of AI tools represents a fundamental shift from today’s widely used generative AI.

Unlike large language models such as ChatGPT and Claude, AI agents can autonomously gather information, make decisions and produce work with minimal human oversight.

“In the context of grants, generative AI may help you polish or write a better application,” Rees said. “But agentic AI will go off and write the application and submit it for you.”

Such systems can be trained on a researcher’s published work, funding criteria and previously successful grant applications to generate, review and improve proposals.

“The marginal cost of producing an application falls to zero,” Rees said. “That’s important because it changes the nature of the problem. It’s not about polishing your grant. It’s about a system that’s designed fundamentally to assess human judgment and human ideas being fed outputs that simulate human judgment.”

Rees and James Wilsdon, a professor of research policy at UCL and executive director of the Research on Research Institute, said AI agents are creating three major challenges for research funders—a massive increase in applications, “quality compression” and “convergence.”

The pair’s recent research analyzed data from 12 funders across seven research systems and found grant application volumes had increased by 57 percent between the launch of ChatGPT in 2022 and the end of 2025. Early data for 2026 suggests the rate of growth is accelerating even more, Wilsdon said.

The second challenge is what the researchers described as “quality compression.”

“When everyone’s proposal has optimization and great writing, the quality floor rises, but the ceiling stays the same,” Rees said. “It becomes harder to discriminate between really excellent ideas and ones that are just OK.”

The longer-term concern is convergence, where AI systems generate grant applications while also assisting reviewers in assessing them.

“If you’ve got agents writing, and agents reviewing, and the agents are talking to each other, they’re all trained on exactly the same work,” Rees said. “Over time, the system isn’t going to evaluate great ideas. It’s just going to measure how well agents simulate what funders have previously rewarded. That’s a real existential threat to what we’re trying to do.”

The researchers warned that the speed of technological change is likely to outpace the ability of universities and funders to adapt.

“If you think it’s bad now, just wait another year,” Rees said. “The tools are improving at such a rapid rate that we run the risk they’ll outpace our ability to transform the grant-funding system.”

Both researchers rejected suggestions to ban AI in grant writing, saying it is impossible to police. 

“Bans are not enforceable,” Rees said. “Attempts to detect AI produce a huge number of false positives. It’s not a practical approach.”

The pair argued that funders and universities should work together to redesign research assessment.

“Currently funders are bringing short-term fixes that often offload some of the work to universities,” Rees said. “But it’s about thinking deeply about the architecture of the current system, which is no longer fit for purpose because of LLMs and agentic AI.

“We need to think about evaluating based on what agentic AI can’t simulate: research track record as appropriate to a researcher’s career stage … track record of delivering great ideas and thinking deeply about what a great idea is in the first place. So that’s going to mean a different emphasis in how we evaluate people and grants.”