"Could agentic AI topple grant-funding systems?" is a Comment by Geraint Rees and James Wilsdon published in Nature, Vol 652, 30 April 2026, p.1119, in the journal's "Setting the agenda in research" Comment section. It argues that AI agents are now capable of producing tens of high-quality grant applications in minutes, and that funders must act before the resulting volume of polished proposals makes it impossible to distinguish strong applications from weak ones.
Authors: Geraint Rees, James Wilsdon Publication: Nature, Vol 652, 30 April 2026, p.1119 — Comment section ("Setting the agenda in research")
Rees is a UCL neuroscientist and former pro-vice-provost; Wilsdon is a science-policy professor at UCL. Both are UK science-policy voices.
Summary of argument
Rees and Wilsdon describe AI agents as LLMs equipped with tools that let them search the web, read documents, write and execute code, and call external services. They state that such agents can be trained on three inputs: a researcher's entire published body of work; the grant criteria of the most relevant funding panel; and the texts of the most recently funded grants from that panel, which they note are often publicly available. From these, the authors write, an agent can produce tens of fully-formatted grant applications "in minutes, and with little work by the researcher." They present this as a description of currently available tools rather than a forecast.
The central claim is what the authors frame as a wheat-from-chaff problem, quoting their own formulation: "Funders must take action before an increase in high-quality proposals written using AI models makes it impossible to sort the wheat from the chaff." The authors argue that funding panels' review capacity scales linearly while AI-assisted application volume can scale faster, and that reviewers cannot reliably distinguish AI-assisted high-quality applications from human-authored ones because both produce the same surface artifact. On this account the prose quality of an application ceases to function as a signal of underlying quality.
Key claims
The authors set out several specific ways they argue the dynamic strains grant-review systems:
- Panel signal. If a large share of applications, illustrated with a figure of 80%, are AI-assisted and equally polished, panels can no longer use prose quality as a quality signal.
- Panel throughput. If application counts triple, review burden triples while reviewer hours remain fixed.
- Reviewer trust. If applications are routinely AI-generated but unattributed, the reviewers' assumption that they are reading the researcher's authentic thinking breaks down.
- Research-priorities fidelity. Agents trained on what has recently been funded implicitly bias toward continuity rather than novelty, which the authors describe as a drift away from a funding agency's stated portfolio strategy.
The authors do not provide quantitative metrics for the rate or quality of agent-generated applications; the illustrative figures (80% of applications, a tripling of counts) are used to frame the argument rather than reported as measured data.
Provenance and confidence
The piece is a short Nature Comment treated here as Rees and Wilsdon's position rather than as established fact. Its factual premise — that current agentic AI can produce grant applications "in minutes, and with little work by the researcher" — is verifiable in principle, but the article supplies no quantitative measurements; such figures would need to come from independent analyses. A future iteration of the Anthropic Economic Index tracking grant-writing tasks could quantify the scale of the disruption directly; the Anthropic Economic Index — March 2026: Learning Curves release does not yet break out research-grant-writing as a tracked task. Confidence is rated medium.
Relationships
- supports: Agentic AI (domain-specific case study of agentic-AI institutional disruption), Ai And Academia (planned; grant-funding disruption as one strand of broader AI-academia disruption, alongside plagiarism and peer-review fakery)
- related: AI as Normal Technology (the authors frame routine deployment of currently-available agentic AI as breaking an existing institution), Anthropic Economic Index — March 2026: Learning Curves (could quantify the scale), Anthropic Economic Index — January 2026: Economic Primitives, Geraint Rees (planned), James Wilsdon (planned), AI Coding Agents, AI Agentic Browsers
- instance-of: Agentic Ai Institutional Disruption (planned umbrella concept)
Future industry coverage could add research funding as a deployment domain (the industries/ folder).