Verdict: this tiny Nature correspondence absolutely delivers, because it names the boring-but-deadly problem hiding under the shiny AI panic: too many proposals, not enough human attention.

That is the whole mess in one sentence. Davies, Bann, Keller, and Munafò are responding to a recent Nature Comment by Geraint Rees and James Wilsdon, who warned that agentic AI could swamp grant-funding systems by generating polished applications at scale. Not "help me rewrite this aim so it sounds less like a tax form." More like: find the call, draft the proposal, critique it, revise it, maybe submit it while the scientist is making coffee and questioning their life choices.

The Davies team says: fair, that sounds bad. But the deeper issue is proposal volume. AI is not inventing the traffic jam. It is handing everyone a faster car.

The Weird Genius Of A Proposal Pollution Market

Their suggestion is delightfully strange: what if research funders used something like cap-and-trade?

In climate policy, cap-and-trade means you set a hard limit on emissions, issue credits, and let organizations trade those credits. The point is not that carbon markets are magical. They are not. Carbon markets have enough fine print to stun a lawyer. But the core idea is simple: if a shared system has limited capacity, stop pretending capacity is infinite.

Now swap carbon for grant proposals.

A university, department, or researcher would get a limited number of submission credits. Want to submit more? You need more credits. Submit fewer? Maybe you can trade or bank them. Suddenly, proposals have an opportunity cost. Not a fake one, like "please be mindful of reviewers' time," which has the enforcement power of a Post-it note on a blender. A real one.

That sounds harsh until you remember the current system already has costs. They are just hidden. Researchers spend months writing applications that often go nowhere. Reviewers donate nights and weekends. Funders triage piles of documents written in the sacred dialect of "broader impacts, but make it grant-ish." AI makes that pile taller.

Why AI Makes This Worse

Large language models are neural networks trained on huge amounts of text to predict and generate language. AI agents build on that by chaining tasks together: search, draft, revise, check forms, repeat. Your autocomplete went to business school and came back with a clipboard.

A 2026 arXiv study by Qian and colleagues found that LLM use in US federal funding appears to have risen sharply after 2023, and that more AI-involved proposals tended to be semantically closer to recently funded work. Translation: AI may help proposals sound more like the proposals that already win. That is useful if you want funding. Less cute if you want weird, risky, field-bending science instead of 400 proposals all wearing the same blazer.

Nature also reported that AI-assisted NIH proposals were more likely to win funding, but might nudge research toward safer, less distinctive ideas. Honestly? That is the most grant-system outcome imaginable. We built a machine that can imitate successful grant language, then acted surprised when it imitated successful grant language.

And researchers are already using these tools. Liao and colleagues surveyed 816 verified researchers and found that 81% had incorporated LLMs somewhere in their workflow. The genie is not going back in the bottle. The genie has a Zotero library now.

The Fairness Problem Sitting In The Corner

The obvious worry: proposal caps could punish early-career researchers, people without famous mentors, smaller institutions, or anyone not already plugged into the funding machine. If credits get hoarded by rich universities, congratulations, we reinvented academic Monopoly and somehow made it less fun.

That is why the cap-and-trade version matters more than a blunt "six proposals per person" rule. A thoughtful credit system could reserve credits for new investigators, protect smaller institutions, or weight credits by grant size. Big center grant? Costs more credits. Small exploratory proposal? Costs fewer. Think of it as surge pricing, except instead of Uber taking your money during rain, it is the scientific ecosystem trying not to eat itself.

Still, this would need careful design. Markets can reveal priorities, but they can also reward gaming. If the credits become another administrative ritual, researchers will develop credit strategies, credit committees, credit consultants, and eventually a webinar called "Ten Secrets To Optimizing Your Proposal Credit Portfolio." Nobody wants that. Nobody emotionally survives that.

What This Could Actually Fix

The best version of this idea changes the question from "How do we detect AI-written grants?" to "How do we make the system robust when grant-writing gets cheaper?"

That is smarter. AI detection is brittle, unfair, and kind of absurd in a genre where everyone already writes like a committee trapped inside a thesaurus. A proposal cap attacks the incentive problem instead. If submissions are limited, researchers may spend more time choosing the right ideas, funders may see fewer low-effort applications, and reviewers may finally read proposals without needing the thousand-yard stare of a person who has scored 37 "innovative" projects before lunch.

There is also a document angle here. Grant workflows are drowning in PDFs, call documents, budgets, biosketches, and forms. Tools like pdfb2.io make browser-based PDF handling less painful, which is nice, because apparently modern science requires both intellectual courage and the ability to merge attachments without screaming.

The Bottom Line

This correspondence is not a full policy blueprint. It is more like someone walking into a smoky kitchen and saying, "Maybe the toaster should not be on fire." Useful.

The cap-and-trade idea will not solve every problem in grant funding. It will not make peer review perfectly fair. It will not stop AI from producing sentences that sound confident while quietly wearing clown shoes. But it does shift attention to the right place: scarcity.

Reviewer time is scarce. Funding is scarce. Attention is scarce. AI makes writing cheaper, but it does not make evaluation free.

That is the part worth sitting with. If proposal generation becomes nearly effortless, the funding system cannot keep pretending every application costs the same to society as it does to the submitter. The bill arrives somewhere. Usually on a reviewer's Saturday.

References

• Davies, N. M., Bann, D., Keller, M., & Munafò, M. (2026). "Is it time to 'cap and trade' credits for research-funding proposals?" Nature, 654, 290. DOI: 10.1038/d41586-026-01777-1. PMID: 42230834
• Rees, G., & Wilsdon, J. (2026). "Could agentic AI topple grant-funding systems?" Nature, 652, 1119-1121. DOI: 10.1038/d41586-026-01297-y
• Qian, Y. et al. (2026). "The Rise of Large Language Models and the Direction and Impact of US Federal Research Funding." arXiv: 2601.15485
• Liao, Z. et al. (2024). "LLMs as Research Tools: A Large Scale Survey of Researchers' Usage and Perceptions." arXiv: 2411.05025
• Liang, W. et al. (2025). "Quantifying large language model usage in scientific papers." Nature Human Behaviour, 9, 2599-2609. DOI: 10.1038/s41562-025-02273-8

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.