The Algorithmic Crisis: How Agentic AI Threatens the Foundations of Global Research Funding

The global research ecosystem is currently navigating a period of unprecedented volatility. For decades, the process of awarding research grants—the lifeblood of scientific innovation—has relied on a delicate balance of human expertise, peer review, and the subjective evaluation of merit. However, the emergence of "agentic" artificial intelligence (AI) has introduced a disruptive force that experts warn could constitute an "existential threat" to the integrity of this entire framework.

While generative AI tools like ChatGPT have already begun to reshape how academics write, the next frontier—autonomous AI agents—represents a paradigm shift. Unlike their predecessors, these agents can function with minimal human intervention, effectively conducting independent research, synthesizing literature, and generating full-scale grant proposals. As these tools proliferate, the mechanisms designed to reward human brilliance are beginning to buckle under the weight of synthetic volume and algorithmic conformity.

The Technological Shift: From Assistants to Agents

To understand the scope of the crisis, one must distinguish between the generative AI of the recent past and the autonomous agents of the present. Generative AI is, at its core, a sophisticated tool for drafting and refinement; it is a "copilot" that requires a human pilot to guide its direction and provide the foundational intent.

Agentic AI, by contrast, operates on a different plane. These systems can be tasked with high-level objectives—such as "write a successful proposal for this specific call for funding"—and proceed to execute the necessary sub-tasks autonomously. By training these agents on a researcher’s previous publications, the specific requirements of funding bodies, and historical datasets of successful grant applications, these systems can produce work that is virtually indistinguishable from human-authored proposals.

Geraint Rees, vice provost for research, innovation and global engagement at University College London (UCL), notes that this shift fundamentally alters the economic landscape of research. "In the context of grants, generative AI may help you polish or write a better application," Rees explained during a recent webinar hosted by the League of European Research Universities. "But agentic AI will go off and write the application and submit it for you."

Chronology of a Disruption: The Post-ChatGPT Explosion

The trajectory of this disruption has been rapid, moving from theoretical concern to empirical reality in a matter of months.

• Late 2022: The public release of ChatGPT triggers an immediate, though largely manual, adoption of LLMs in academia. Researchers begin using AI to summarize literature and assist in the drafting of grant sections.
• 2023–2024: The "polishing" phase. Funders begin to notice an increase in the homogeneity of writing styles, though the volume of applications remains within manageable, albeit growing, parameters.
• 2025: The rise of agentic frameworks. AI agents move beyond mere writing to project management, data synthesis, and autonomous proposal generation.
• 2026 and Beyond: The "acceleration phase." Early data from the current year indicates that the growth rate of applications is outpacing all previous historical trends, leading to systemic strain on peer-review infrastructure.

The impact of this timeline is stark. Research conducted by Rees and James Wilsdon, a professor of research policy at UCL and executive director of the Research on Research Institute (RoRI), highlights a 57 percent increase in grant application volumes between the launch of ChatGPT in 2022 and the end of 2025. Data from 2026 suggests this rate is not merely continuing but accelerating, threatening to overwhelm the administrative capacity of funding institutions globally.

The Triad of Challenges: Volume, Compression, and Convergence

Wilsdon and Rees have identified three specific, interlocking crises that define the threat to the research funding architecture:

1. The Volume Crisis

As the marginal cost of producing a grant application approaches zero, the incentive structure for researchers shifts. If an agent can generate a perfectly formatted, compliant proposal in seconds, the barrier to entry for grant competitions effectively vanishes. This leads to a "flood" of applications that creates a bottleneck for human reviewers, who are already in short supply.

2. Quality Compression

"When everyone’s proposal has optimization and great writing, the quality floor rises, but the ceiling stays the same," Rees noted. This phenomenon, termed "quality compression," makes the task of discrimination nearly impossible. When every application is technically proficient, grammatically flawless, and perfectly aligned with the funder’s criteria, the human reviewer is stripped of the ability to distinguish between "really excellent" ideas and those that are merely competent.

3. Algorithmic Convergence

Perhaps the most insidious danger is "convergence." This occurs when the systems writing the grants and the systems assisting in the review of those grants become interconnected. If both sides of the equation rely on AI trained on the same historical datasets, the process loses its capacity for novelty. Instead of identifying breakthrough research, the system begins to reward applications that best mimic the patterns of the past. As Rees warns, "The system isn’t going to evaluate great ideas. It’s just going to measure how well agents simulate what funders have previously rewarded."

Official Perspectives and the Failure of Policing

In the face of this systemic threat, there has been a reflexive impulse among some institutions to ban the use of AI in grant writing. However, experts like Rees and Wilsdon have categorically rejected this approach as both impractical and unenforceable.

"Bans are not enforceable," Rees argued. "Attempts to detect AI produce a huge number of false positives. It’s not a practical approach." The cat-and-mouse game of AI detection tools versus AI-generated content is a losing battle for universities, as the technology to generate text consistently evolves faster than the tools designed to flag it.

Instead, the consensus among policy experts is shifting toward systemic reform. The current architecture of research assessment—which prioritizes standardized, written proposals—is increasingly seen as "no longer fit for purpose."

Implications: Redesigning the Architecture of Merit

The "existential threat" identified by the researchers is not necessarily to the research itself, but to the gatekeeping mechanisms that define what research gets funded. To survive the era of agentic AI, the global research community must move toward a model of assessment that prioritizes qualities that AI cannot simulate.

Moving Beyond the Proposal

If the proposal itself can be perfectly synthesized, it can no longer be the primary metric for judging merit. Experts suggest that funders should shift their focus toward:

• Track Record Evaluation: Emphasizing a researcher’s long-term career trajectory, their history of delivering impactful work, and their demonstrated ability to navigate complex research challenges.
• Human-Centric Review: Re-introducing intensive, human-to-human interview processes or oral defenses, where the depth of a researcher’s understanding can be tested in real-time, away from the influence of generative agents.
• Collaboration and Transparency: Universities and funders must break down the current silos, sharing data on application trends to create a more resilient, adaptive funding architecture.

Conclusion: A Race Against Time

The speed of technological change is currently outstripping the ability of the academic establishment to adapt. As Rees put it, "If you think it’s bad now, just wait another year." The risk is that if funding bodies fail to evolve, the entire system of research support will descend into a cycle of synthetic applications and synthetic evaluations, effectively hollowing out the creative, human-led inquiry that drives scientific progress.

The call to action is clear: the architecture of research funding must be fundamentally re-engineered. The focus must shift from the "what" of the grant proposal—which is now easily commodified—to the "who" of the researcher. Failure to pivot in this direction risks turning the pursuit of knowledge into an automated feedback loop, where the only thing being tested is the proficiency of the algorithms we have unleashed.