AI for research proposal writing: what it can and cannot do in 2026

Let us narrow the scope first: this guide is for researchers preparing research proposals and funding applications (an MSCA fellowship, an ERC grant, a national project, a fellowship application). Most searches on this topic ask for a generator: a tool that writes the proposal for you. That is the wrong question. Producing text has never been cheaper; what remains expensive is what always was: a solid idea, a track record that backs it, and the ability to answer for every claim in front of a panel. Well-used AI accelerates the first without compromising the last.

Where generative AI genuinely helps

Used with judgement, generative AI is genuinely useful in specific parts of an application:

• Structuring against the call criteria: decomposing the call text into requirements and checking that each one is answered somewhere in your draft.
• Adapting your existing material: re-projecting your CV, publications and merits into the format and emphasis each call demands, instead of rewriting everything from scratch every time.
• First drafts of standard sections: dissemination plans, data management, timeline structure. Always as a proposal to review, never as final text.
• Editing and language: clarity, concision, academic English. The lowest-risk, highest-immediate-return use.
• Consistency review: spotting contradictions between sections, undefined acronyms, promises in the abstract the work plan does not keep.

What it cannot do (and what funders require)

A generative AI does not know your track record, cannot tell a real citation from a perfectly formatted invented one, and cannot be an author of anything. Funders have now put this in writing. The Horizon Europe standard application form permits AI use but makes you fully responsible for the output, requires transparency about the tools used and a list of sources; failing to comply with those guidelines can make a proposal ineligible. The Commission's ERA Living Guidelines, revised in May 2026, set the integrity framework. National systems are following: in Spain, ANECA criteria require declaring generative AI use whenever it affects the original content of a contribution.

We cover the full rules, with sources, in Can you use AI in your EU grant proposal? What Brussels requires in 2026.

The generator trap

Proposal generators promise the whole project from a paragraph. Three practical problems:

1. Generic text: panels read dozens of applications; template prose is recognisable, and with success rates compressed by the flood of AI-assisted submissions, generic competes worse than ever.
2. Invented citations and data: the generator does not verify; you answer for it. A single non-existent reference spotted by an evaluator damages the credibility of the whole proposal.
3. Confidentiality: pasting your unpublished idea, or collaborators' material, into a consumer tool can breach confidentiality guidance and compromise your own scientific priority. The ERA Living Guidelines flag this explicitly.

A workflow that survives scrutiny

The workflow that works inverts the generator's order: verified facts first, text second.

1. Consolidate your record once, with every merit linked to its source document: publications, contracts, research stays, teaching, certificates.
2. Decompose the call into requirements and map which part of your record answers each one. The gaps that appear are information, not something to paper over.
3. Let AI propose from your confirmed data, never from a blank page: every claim must originate in a fact of yours.
4. Confirm every block yourself before it enters the draft. Human review is not a final formality, it is the control mechanism.
5. Log and disclose: which tools you used, for what, and with which sources. It is what the European form can require and what national criteria increasingly demand.

How to choose a tool

More useful than a list of names, which goes stale every quarter, is a list of criteria. Before adopting any AI tool for applications, check:

• Traceability: can every claim be traced to a fact of yours or a source? Or does text simply appear with no origin?
• Who confirms: does the tool propose and you approve, or does it ship whatever it generates?
• European format depth: does it understand MSCA, ERC, national CV formats such as Spain's CVN and CVA, or is it built only for US federal calls?
• Data protection: where is your data hosted, does GDPR apply, are your texts used to train models?
• Disclosure support: does it help you record what was AI-generated and from which sources, to meet the transparency the forms require?

Any tool that fails on traceability or on human confirmation is asking you to carry a risk that funders have already put in writing is yours.