Expert insight by Annelot Staes • 5 min read

In January, I wrote about what a Nature column called a “point of no return” in grant writing: the moment when the combined effort researchers put into applying starts to outweigh what the funding brings back to the research community. [1] In some Horizon Europe calls, the time invested by all applicants together can exceed the budget of the call itself. For researchers, that means months of work on proposals that are most likely to be rejected, time that is then not spent on research.

Seen in that light, it is perfectly logical that researchers are turning to AI. If a tool can cut a proposal from weeks to days, and the odds of success are low anyway, why wouldn’t you use it? The recent figures from the European Research Council on Starting Grant applications for the 2027 call show the new set of problems we will be running into in the coming year.

Record demand, record-low odds

The ERC is being swamped. For the 2027 Starting Grant call, the ERC scheme for early-career researchers that closed on 14 October, applications rose by about 60% compared to last year: from just over 4,800 to more than 7,700. At the same time, the number of available grants fell, from 421 to at most 373. Logically, this has caused the success rate to drop further. It was about 12% in the 2025 call, 8.8% in 2026, and is expected to fall below 5% for 2027. [2,3]

Review panels under pressure

Besides that, a practical problem is arising: the number of proposals that ERC review panels need to get through is rising fast. ERC President Maria Leptin wrote in April that some panels now have more than 250 applications to assess, where 50 to 150 used to be normal, and that the ERC expects this trend to continue. [4] Reviewers have a limited amount of time, and an ERC review is meant to be deep and careful. Something has to give.

In an attempt to bring down the number of applications, the ERC announced in April that existing waiting periods for unsuccessful applicants to resubmit would be extended by one year. More than a thousand researchers signed an open letter in protest, and within two weeks the ERC withdrew this change for the Starting, Consolidator, Advanced and ERC Plus Grants. Only the Synergy Grants kept the stricter rules. [4,5,6]  Nonetheless, it shows that the system is under real strain: the ERC itself says that different measures, such as changes to the resubmission rules, may follow.

A Europe-wide trend since ChatGPT

And it is not just the ERC. If you look across Europe, the same pattern appears in other schemes. Researchers at UCL and the Research on Research Institute analysed application data from 12 funders in six countries and found that application volumes in 2025 were on average 57% higher than in 2022, the year ChatGPT launched. [7] In the US, the NIH has even limited principal investigators to six applications per year. [8] So the issue is not one scheme with a design problem. In Europe and beyond, excellent ideas have become abundant, and money has not. Funders are experimenting with restrictions on resubmission, caps on applications and even partly drawing lots among top-ranked proposals.

When everyone writes well, nobody stands out

Using AI makes sense for each researcher individually. The problem appears when everyone does it. Writing becomes cheap, so more people submit. And the average text gets better, so quality stops separating applicants. Lynn Kamerlin, writing in EMBO Reports, tested a chatbot on the knowledge-transfer section of an MSCA postdoctoral fellowship application. [3] The result was convincing, and she concludes that a shrinking share of each proposal is truly individual. Research on US NIH and NSF proposals also suggests that AI-assisted proposals tend to resemble previously funded ones. [9]

This matches what I see when supporting and reviewing ERC grants. More texts look professional, but fewer feel like they could only have been written by that one researcher. In other words, the time saved by AI has not made winning easier for researchers as a group. It has raised the baseline everyone has to meet.

What AI can’t do for you

I see four things AI cannot do for you:

1. Tell you whether the text is right. An AI draft can sound convincing and still miss the point. It may promise more than the budget allows, contradict another section, or ignore what the scheme actually asks for. Checking a text against the rules and the review criteria takes experience with how the scheme works in practice.
2. Make your idea unmistakably yours. AI is very good at producing what a good proposal usually sounds like. That is exactly the problem. If everyone’s text converges on the same average, the proposals that stand out are the ones with a sharp question, a clearly stated risk, and a personal reason why this researcher should do this project. That comes from your thinking, not from a prompt.
3. Help the reviewer. A reviewer with multiple proposals to read will not give you the benefit of the doubt if the main message is hard to find, or if they cannot see how your claims connect to your work plan and budget. Clear structure, visual elements and traceability are not decoration. They make the reviewer’s job easier.
4. Know when not to apply, or when to wait. AI makes it easy to submit, but a submission is not free of consequences. In the ERC, a weak result can close the door for a while: under the current rules, a C at the first step blocks the next two calls and a B the next one. And for the Starting and Consolidator Grants you only have a limited window after your PhD. An application sent in “just to see” can therefore cost you the chance to come back next year with a much stronger proposal. A ‘C’ is given when reviewers don’t see enough original, ambitious science, and a polished AI-generated draft can’t fix that. The following considerations apply before you start writing: Is the idea mature and different enough to compete at the first step? Is this the right scheme, and the right moment in your career?

Where the difference is still made

No grant advisor can take the luck out of a process with a success rate under 5%. AI is here to stay, and I use it too. Used well, it saves time on the routine parts. But when everyone has the same tool, the difference is made elsewhere: in judgment, in originality and in care for the reader. That is where I think good grant advice earns its place.

Planning an ERC application? Get in touch for a critical read of your proposal. Contact us at info@hezelburcht.com, call 088 495 20 00 or complete the contact form below:

Contact form

Read my earlier article: The cost of competition in Horizon Europe calls

References
1. Schweiger, G. “Point of no returns: researchers are crossing a threshold in the fight for funding.” Nature, 18 December 2025. https://www.nature.com/articles/d41586-025-04060-x
2. O’Callaghan, J. “Is research funding now a lottery? Elite funder ERC swamped by applications.” Nature 659, 17–18 (28 October 2026). https://www.nature.com/articles/d41586-026-03476-3
3. Kamerlin, S. C. L. “Hypercompetition, score compression and AI co-authorship: a perfect storm for grant success rates.” EMBO Reports 27, 2145–2146 (2026). https://doi.org/10.1038/s44319-026-00779-4
4. Leptin, M. “ERC President explains stricter application measures amid rising demand for funding.” ERC, 16 April 2026. https://erc.europa.eu/news-events/news/erc-president-explains-stricter-application-measures-amid-rising-demand-funding
5. “More than 1,000 scientists urge ERC to reverse resubmission rule.” Times Higher Education, 28 April 2026. https://timeshighereducation.com/news/more-1000-scientists-urge-erc-reverse-resubmission-rule
6. “ERC Scientific Council readjusts rules for reapplication.” ERC, 29 April 2026. https://erc.europa.eu/news-events/news/erc-scientific-council-readjusts-rules-reapplication
7. Rees, G. & Wilsdon, J. “Could agentic AI topple grant-funding systems?” Nature 652, 1119–1121 (2026). https://www.nature.com/articles/d41586-026-01297-y
8. NIH. “Supporting Fairness and Originality in NIH Research Applications,” NOT-OD-25-132 (17 July 2025). https://www.grants.nih.gov/grants/guide/notice-files/NOT-OD-25-132.html
9. Qian, Y. et al. “The Rise of Large Language Models and the Direction and Impact of US Federal Research Funding.” arXiv:2601.15485 (2026). https://arxiv.org/html/2601.15485v1