Can you use AI to write a grant proposal? What NIH, NSF, ERC and UKRI allow in 2026

Yes: every major funder checked for this post lets you use ChatGPT, Claude, Gemini or similar tools to help write a grant proposal, on its terms. The ideas must be yours, you answer for every sentence and reference, nothing may be fabricated or copied, and many funders want the use declared. NIH draws the hardest line: it will not treat applications "substantially developed by AI" as the applicants' original ideas. Reviewers face stricter rules. Almost every funder forbids putting an application into a public AI tool, and the few that allow more allow it only through secure tools, for supporting tasks, never for the judgement itself.

The newer question is whether funders will use AI on proposals themselves. On 3 November 2026, Nature Index published "Will AI grant reviews make science more conservative?" by Elliot Yates.¹ It reports a study in which GPT-4o, choosing between pairs of real proposals to the Austrian Science Fund, agreed with the human funding decision in 60% of cases but was "systematically less willing" than the reviewers to fund proposals rated highly for novelty. The article notes that UK Research and Innovation (UKRI) is exploring AI-assisted assessment. Germany's DFG already lets AI help its staff check proposals, select reviewers and summarise reviews, and the US agency ARPA-H said it would pilot language models in the first pass over one programme's submissions.

The rules have moved fast. Between December 2025 and October 2026, DFG, the Swiss National Science Foundation (SNSF), the ERC, Australia's ARC and NHMRC, and UKRI all rewrote what reviewers may do. Advice from 2024 is often out of date.

Below: what applicants may do and when they must say so; a table of the rules at a dozen funders (NIH, NSF, UKRI, Wellcome, ERC, Horizon Europe, DFG, SNSF, ARC, NHMRC, Canada's federal agencies and TÜBİTAK); what reviewers may do; which funders use AI themselves and what the novelty study found; and a checklist with a disclosure template. Every source was checked on 5 November 2026.

Can I use ChatGPT, Claude or Gemini to write a grant proposal?#

Yes, with conditions. None of the funders below bans AI help for applicants. The clearest summary of the shared conditions is the joint statement of seven UK funders² (AMRC, the British Heart Foundation, Cancer Research UK, the Royal Academy of Engineering, the Royal Society, UKRI and Wellcome), revised in June 2026. It names "ChatGPT, Co-pilot, Claude or Gemini" and lists acceptable uses, from translation and better English to "assisting idea generation or development of hypotheses", summarising data, comparing literature, writing code and generating an abstract. Applicants must not use the tools to fabricate or alter data, plagiarise, fake references, or "to generate an entire application, or sections of an application, without human involvement".

Funders differ on how much of the writing a tool may do:

• NIH allows AI help "for limited aspects or in specific circumstances", but "NIH will not consider applications that are either substantially developed by AI, or contain sections substantially developed by AI, to be original ideas of applicants" (NOT-OD-25-132³, 17 July 2025). AI use found after an award can be referred to the Office of Research Integrity and cost the grant. Because AI had let some investigators submit "more than 40 distinct applications in a single application submission round", NIH has accepted at most six applications per principal investigator per calendar year since 25 September 2025, with a few exceptions. The rule is now part of the March 2026 Grants Policy Statement⁴. A reminder of 14 May 2026 lists "Completely generating grant applications and/or papers using AI tools (including agents)" among possible misconduct, and adds that NIH does not "plan to raise concerns when AI is used in accordance with responsible research practices" (NIH Extramural Nexus⁵).
• NSF holds the proposer responsible for "the accuracy and authenticity" of the proposal, AI-assisted content included.⁶ Since 8 December 2025 its definition of research misconduct covers fabrication, falsification or plagiarism committed "through the use or assistance of other persons, entities, or tools, including artificial intelligence (AI)-based tools" (PAPPG 24-1, Supplement 1⁷).
• Horizon Europe's application form: "Applicants are fully responsible for the content of the proposal (even those parts produced by the AI tool)". They must check every AI-generated citation and list their sources.⁸
• Canada's federal agencies answer "Yes" to drafting or summarising sections with AI: "The accountability for the integrity and quality of the final application remains solely with you, the applicant."⁹
• TÜBİTAK allows AI first drafts of sections but calls them high-risk: never final text, and to be rewritten from the applicant's own knowledge and analysis.¹⁰

Do I have to disclose AI use in a grant application?#

It depends on the funder. The answers fall into four groups:

• Required. Horizon Europe applicants "must be transparent in disclosing which AI tools were used and how they were utilized".⁸ TÜBİTAK requires a declaration of significant use in a dedicated section of its online application, naming the tool and version, the stage or section, and the nature of the use.¹⁰ Wellcome requires substantive use to be declared on the form (Wellcome policy, updated 8 April 2026¹¹). DFG asks "which generative models were used, for what purpose and to what extent" (DFG¹²).
• Expected or advised. UKRI: "applicants and applications are expected to be transparent where they have used generative AI tools" (UKRI policy, updated 6 July 2026¹³). NIH's May 2026 reminder advises researchers to "Clearly describe in applications, manuscripts, and presentations the use of the AI tools".⁵
• Encouraged. NSF proposers "are encouraged to indicate in the project description the extent to which, if any, generative AI technology was used and how it was used".⁶
• Only when asked, or not at all. The ARC expects transparency "when this information is requested by the ARC".¹⁴ Canada: "No, there is no requirement for applicants to disclose the use of generative AI in their applications", though sources that AI surfaced must still be cited.⁹

Minor help usually needs no statement. UKRI and Wellcome count translation into English, better English and formatting or cutting words as minimal, and idea generation, hypotheses, summaries, literature comparison, code and abstracts as substantive. DFG exempts grammar, style and spelling checks and translation tools.¹² TÜBİTAK's bar is lower: anything beyond basic grammar or spelling checks counts, including drafting any section, code, data analysis or images.¹⁰

Will saying so count against you? Four funders say no in writing. UKRI: "This information will not affect the assessment process."¹³ Wellcome collects it "for monitoring purposes" and it "will not be used in the assessment process".¹¹ DFG: AI use "does not affect how a proposal is assessed".¹² Horizon Europe evaluators are told AI use in drafting "may not be considered by expert evaluators as a reason to penalise a proposal".¹⁵ UKRI also tells assessors not to "speculate within their assessment whether generative AI has been used".

For what a good statement contains, see how to write an AI disclosure statement. Grant forms usually give a box and a character limit; a short template follows in the checklist.

What do the major funders allow? Rules compared#

Each row comes from the funder's own pages, opened on 5 November 2026. "Not stated" means the pages checked do not address the point.

Funder (policy date)	Applicants	Disclosure	Reviewers and panels	Funder's own use of AI
NIH (notices of 23 June 2023 and 17 July 2025; reminder of 14 May 2026)	Limited help; work "substantially developed by AI" not original; six applications per PI a year	Advised	No generative AI for critiques; nothing from an application into online AI tools16	AI detection on applications
NSF (notice of 14 December 2023, updated 21 May 2026)	Allowed; misconduct rules cover AI tools	Encouraged	No proposal content in "non-approved" tools; public information allowed	Not stated
UKRI (policy of September 2024, updated 6 July 2026)	Allowed; no application or sections "without human involvement"; no AI in interviews	Expected for substantive use	Language refinement only, with no application content entered	Retrospective trial of AI-assisted assessment announced 10 September 202617
Wellcome (updated 8 April 2026)	Allowed; limits as in the UK joint statement	Required for substantive use	Never input applications; "side tasks" only	None for quality or funding decisions; protected tools for eligibility and remit checks
ERC (Scientific Council position, updated 25 March 2026; reviewer guidelines of 23 March 2026)	Allowed; "full and sole authorship responsibilities"18	Not stated	Language polishing and topic searches; no summaries, merit judgements or draft reviews, even on a local model19	Not stated
Horizon Europe (application form of 22 January 2026; expert briefing of 1 December 2025)	Allowed; applicants "fully responsible"	Required: which tools, and how	No delegation; side tasks only, if confidentiality is ensured	Not stated
DFG (review guideline of March 2026; 2023 review ban lifted December 2025)	Generally permitted	Required: models, purpose, extent	Supporting tasks through local, institutional or contract-protected tools; declared20	Formal checks, reviewer selection, review summaries, funding recommendations, all checked by staff
SNSF (applicant note of 28 February 2024; reviewer guidelines of October 2026)	Allowed; applicants are "sole authors"21	Not stated	After forming their own view: literature, notes into a draft, language; secure tools; declared22	"uses AI applications to help process funding applications"
ARC and NHMRC (policies effective 28 April 2026)	Allowed as a support tool	ARC: when asked; NHMRC: not stated	ARC: grammar, spelling, formatting, readability; NHMRC: refining comments, not scoring23	Not stated
Canada: CIHR, NSERC, SSHRC, CFI (page modified 30 October 2026)	Allowed, including drafting	Not required	Prohibited, even institution-provided tools and draft reviews	Not stated
TÜBİTAK (guide of January 2026)	Allowed; AI-drafted sections high-risk; no confidential or personal data in tools	Required for significant use	Prohibited for any purpose connected with evaluation	Not stated

Two things the table cannot show:

• The call text wins. A call can be stricter than the funder's general policy. UKRI points applicants to the "specific funding opportunity page",¹⁷ the UK joint statement to each funder's own declaration rules, and TÜBİTAK said in October 2025 that it would add its guide to every programme's application page and to reviewers' forms.¹⁰
• Penalties. Where an allegation of misconduct involving generative AI is upheld, UKRI may reject the person's applications and bar them from applying "for any period of time, including indefinitely".¹³ The ARC removes an assessment when inappropriate AI use by the assessor is suspected.¹⁴

May reviewers use AI to assess a grant proposal?#

Not to do the assessment. Beyond that, funders now fall into three groups.

No generative AI in the review at all. Since June 2023, NIH has prohibited reviewers "from using natural language processors, large language models, or other generative Artificial Intelligence (AI) technologies for analyzing and formulating peer review critiques" (NOT-OD-23-149¹⁶); its May 2026 reminder repeats the ban.⁵ Canada's agencies forbid uploading application material even to "generative AI tools procured and deployed by their institution", and even uploading your own draft review (FAQ⁹). TÜBİTAK bans generative AI for any purpose connected with evaluation, including notes for panel meetings and emails about a review (TÜBİTAK guide, January 2026¹⁰).

Language help, with no proposal content in the tool. Until 6 July 2026, UKRI assessors could not use generative AI even "to correct language, spelling, grammar and formatting". Now they may use it "for language refinement purposes", provided no part of the application is entered and the tool is not "tasked with understanding, summarising or evaluating the application's content".¹³ The ERC's guidelines of March 2026 allow polishing language and searching for information on the topic (ERC, 24 March 2026¹⁹). A reviewer may not have even a local, offline model summarise the proposal, may not ask a tool whether the review misses important points, and may not start from an AI draft: "Generating a draft review involves evaluative judgement and constitutes a delegation of the evaluation task, even if the reviewer later revises the text." The ARC¹⁴ allows only grammar, spelling, formatting and readability, and calls putting application material into public tools "such as ChatGPT, Gemini, Claude, Perplexity" a "serious breach of confidentiality". NHMRC²³ allows help with review comments but not using AI "to evaluate, critique and/or score applications". Wellcome, NSF and Horizon Europe allow side tasks that keep the application out of the tool, such as background searches.

Secure tools that may see the proposal, in a supporting role, declared. In 2023 DFG called AI use in reviews "inadmissible".²⁴ That rule has not applied to reviews since December 2025, and DFG now permits AI in preparing reviews under a binding guideline (DFG²⁰). Allowed tools run locally, at a trustworthy institution, or in a commercial cloud under a licence that excludes training and storage. Allowed tasks include literature searches, revising your own draft and "expanding one's own bullet points or notes". Use must be declared in DFG's elan system. "Uncritical adoption of AI-generated text is not permitted." A breach may count as scientific misconduct. The SNSF announced similar rules on 20 August 2026 (SNSF²⁵). Its October 2026 guidelines allow AI only once reviewers "have formed their own opinion", not "to generate all or part of the review" or a list of strengths and weaknesses, and only with "no storage by the provider, no use for training purposes".²²

The common line is non-delegation: the reviewer reads the whole proposal and forms the judgement. Journal peer review has different rules; see what 13 publishers, journals, funders and conferences let reviewers do with AI.

Do funders use AI to assess grant proposals?#

A few do, for steps around the expert judgement rather than the judgement itself:

• DFG's Head Office "may use AI tools to assist in processing funding proposals": formal checks, help selecting reviewers, summarising reviews and preparing funding recommendations, with staff checking every output. "Decisions on funding proposals are not made automatically by AI systems." Applicants accept these conditions through a declaration of rights in DFG's elan system: "It is therefore not possible to object to the responsible use of AI in the review process."¹²
• "la Caixa" Foundation began AI pre-screening of its CaixaResearch Health call in 2023, according to a paper by foundation staff and technical partners (Carbonell Cortés et al., Data & Policy, 2024²⁶). Three classification models trained on past proposals and their review scores flag a proposal only if all three put it in the bottom class, and two human "eligibility reviewers" check every flag; if either has a reasonable doubt, it goes to peer review. In 2023 the models flagged 98 of 493 proposals; 63 were removed and 35 rescued.
• ARPA-H said its IGoR programme, launched on 5 May 2026, "will pilot the use of secure large language model (LLM) tools to assist with the initial review of submitted materials", while "Human experts will remain responsible for all evaluation and decisions" (ARPA-H²⁷).
• UKRI announced on 10 September 2026 eight measures from October, including staff-led desk rejection and "exploring the safe use of AI-assisted assessment", starting with a retrospective trial with the UK Metascience Unit "to compare what AI would have concluded from past applications against the decisions that were made" (UKRI¹⁷). One stated principle is "maintaining the centrality of human judgement to funding decisions".
• Others. The UK joint-statement funders do not use generative AI to make funding decisions but may process applications with "enterprise data protected" tools,² at Wellcome for eligibility or remit checks.¹¹ The SNSF "uses AI applications to help process funding applications" (SNSF, 28 February 2024²¹). NIH uses AI detection to identify AI-generated applications.³

One case shows what happens when applicants are not told. Times Higher Education reported on 25 September 2026 that CRANE, a cybersecurity research network funded through UKRI's Engineering and Physical Sciences Research Council, had used an "AI-based review" to triage 179 proposals; only the top half reached human reviewers (Times Higher Education²⁸). On 5 October it reported that CRANE would re-review every application the AI had screened out and had suspended the offers made. CRANE said it had used "closed, ring-fenced read-only LLMs", and apologised: "we recognise that we were not transparent enough about this approach, and for that we are sorry" (Times Higher Education²⁹).

The fourth version of the European Commission's living guidelines on generative AI in research³⁰, published on 15 October 2026, welcomes such experiments but asks for "high levels of transparency with applicants and reviewers" and for safeguards "particularly in screening, ranking, prioritisation or exclusion". It also warns funders about hidden prompts in proposals, invisible text meant to steer an AI reader "toward a more favourable assessment", and recommends rules against them.

Will AI grant reviews make science more conservative?#

That is the worry the Nature Index piece raises, and the study behind it is small but pointed. Diogo Machado, who heads the data-science team at the consultancy Technopolis, took 139 proposals submitted in 2021 to the Austrian Science Fund (FWF) in mathematics, physics and astronomy, chemistry and biology (Machado, Scientometrics, 2026³¹). Each came with reviewers' scores for novelty, scientific quality, team capacity and feasibility, and the real funding decision. GPT-4o compared proposals in pairs within each field, prompted only with "Which proposal should be funded? Answer with A or B only.", and those with the most wins were selected for funding.¹

The model matched the human decision in 60% of cases. Of the four review scores, novelty was the only one significantly associated with disagreement: "Proposals rated as more novel by human reviewers are substantially more likely to be funded by humans and not by the LLM."³¹ The paper calls this "consistent with a conservative tendency in LLM-based evaluation", and says its design cannot identify the mechanism.

The limits are real: one model, one bare prompt, one funder, four fields, 139 proposals, and no assessment criteria given to the model, a limitation the Nature Index article notes.¹ It adds that a model asked to judge novelty directly might not manage either, citing Susan Guthrie of RAND Europe, who says novelty is highly subjective and hard to standardise. The finding is a warning, not a measurement of any funder's system.

The risk is greatest where funders actually use AI: early triage, where a proposal can be dropped before an expert reads it. That is why la Caixa sends every flag to two people, and why its rescue figure, 35 of 98 flags in 2023, is worth noting.²⁶ For applicants the lesson is modest: write for the expert reader, say plainly what is new and why it matters, and never try to game a machine reader.

What should I do before submitting? A checklist#

For applicants:

• Read the call's own rules first. They can be stricter than the general policy, and the disclosure question may be on the form.
• Keep the ideas yours. Aims, hypotheses and the case for novelty should come from you. Use AI on parts you can check, and rewrite what it drafts.
• Check every reference and factual claim. Several funders name fake references as possible misconduct. See how to check for hallucinated references.
• Check what the tool does with your text. Read the data-use terms before pasting unpublished ideas into a consumer chat app; Canada's FAQ advises reviewing "the AI platform's privacy policies, data retention practices, reuse terms".⁹ Never paste other people's personal data without their consent (UKRI), or confidential or unpublished data (TÜBİTAK).
• Disclose where asked, in the funder's terms. Name the tool and model, what it did and where, and say you checked it.
• Keep a record. Save the prompts, outputs and drafts that reached the application; the disclosure guide has a one-line log format.
• No hidden prompts, no AI in interviews. The Commission now tells funders to prohibit the first;³⁰ UKRI bans the second.

A disclosure sentence to adapt (not any funder's official wording; follow the form's own fields, such as TÜBİTAK's tool, version, stage and nature):

"Generative AI use: [Tool and model, as shown in the interface] ([developer]) was used in [month year] to [summarise literature on X / suggest a structure for section Y / write analysis code for Z / improve the English of the summary]. The research ideas, aims and hypotheses are the applicants' own. We checked every statement and reference against the original sources, revised all AI-assisted text, and take full responsibility for the content."

For minimal use where a funder still asks: "[Tool] was used only to correct grammar and to shorten the text to the word limit. It added no content."

For reviewers:

• Treat the proposal as confidential. Never upload it, or your notes on it, to a public chatbot; at NIH, Canada's agencies and TÜBİTAK, use no generative AI for the review at all.
• Form your own judgement first. Even where some help is allowed, no funder checked lets a tool make the judgement or write the assessment for you.
• Declare it where asked, as DFG, SNSF and the ARC require, and do not guess in your review whether the applicant used AI.

Where does Papex fit?#

A grant disclosure is a sentence in a form; the record behind it is yours to keep. Papex keeps that kind of record for research done with AI. A Papex record states who did what, with CRediT roles, and how AI was involved, as the authors state it; model and tool names are self-reported. The person accountable signs one exact version, and a salted fingerprint of it is timestamped with OpenTimestamps and anchored in Bitcoin, so anyone can check the signer, version and date at /verify. Drafts stay private while you work, and after signing you choose whether a version stays private, is shared or is published. A signed, dated record of a hypothesis or plan can back the AI-use sentence in an application, or show what you had and when (how to protect a research idea).

A record shows who signed which version, when, and what they said the tools did; it is not peer review and does not show that the ideas are correct or original.

Sources#

1. Elliot Yates, Will AI grant reviews make science more conservative? (Nature Index, 3 November 2026), checked 5 November 2026; the opening of the article is free to read and the rest needs a log-in, so only the open part was used ↩ a b c
2. Funders joint statement: use of generative AI tools in funding applications and assessment (Research Funders Policy Group, revised June 2026, published by Wellcome), checked 5 November 2026 ↩ a b
3. NOT-OD-25-132: Supporting Fairness and Originality in NIH Research Applications (NIH, 17 July 2025), checked 5 November 2026 through the Internet Archive copy of 24 October 2026; the live page blocked automated access ↩ a b
4. Summary of Significant Changes to the NIHGPS for March 2026 Version (NIH), checked 5 November 2026 through the Internet Archive copy of 26 August 2026, with NOT-OD-26-057 (NIH, 25 March 2026) (Internet Archive copy of 24 September 2026) ↩
5. Helpful Reminders to Ensure Integrity of NIH-Supported Research When Using Artificial Intelligence (NIH Extramural Nexus, 14 May 2026), checked 5 November 2026 through the Internet Archive copy of 23 July 2026; the live page blocked automated access ↩ a b c
6. Notice to Research Community: Use of Generative Artificial Intelligence Technology in the NSF Merit Review Process (NSF, 14 December 2023, last updated 21 May 2026), checked 5 November 2026 ↩ a b
7. PAPPG 24-1, Supplement 1 (NSF, effective 8 December 2025), checked 5 November 2026 ↩
8. EU Grants: Application form (HE RIA and IA), V5.1 (European Commission, 22 January 2026), Part B guidance on the use of generative AI tools, checked 5 November 2026 ↩ a b
9. Guidance on the use of Artificial Intelligence in the development and review of research funding applications (CIHR, NSERC, SSHRC and CFI, page modified 30 October 2026) and Frequently Asked Questions (same agencies), checked 5 November 2026 ↩ a b c d
10. Destek Süreçlerinde Üretken Yapay Zekânın (ÜYZ) Sorumlu ve Güvenilir Kullanımı Rehberi (TÜBİTAK, January 2026, in Turkish) and TÜBİTAK announcement of the guide (28 October 2025), checked 5 November 2026 ↩ a b c d e
11. Use of Generative Artificial Intelligence (AI) in our grant application process (Wellcome, last updated 8 April 2026), checked 5 November 2026 ↩ a b c
12. AI in Proposal Preparation (DFG), checked 5 November 2026 ↩ a b c d
13. Use of generative artificial intelligence in application preparation and assessment (UKRI, September 2024, last updated 6 July 2026) and its policy page with the list of updates, checked 5 November 2026; the earlier assessor rule from the Internet Archive copy of 29 April 2026 ↩ a b c d
14. Policy on Use of Generative Artificial Intelligence in the ARC's grants programs, version 2026.1 (Australian Research Council, effective 28 April 2026), checked 5 November 2026 through the Internet Archive copy of 16 July 2026; the live site did not respond to automated access ↩ a b c
15. Horizon Europe: Standard briefing for experts, version 13.0 (European Commission, 1 December 2025), checked 5 November 2026 in the copy hosted by Leuphana University Lüneburg; we could not find the Commission's own address for this version ↩
16. NOT-OD-23-149: The Use of Generative Artificial Intelligence Technologies is Prohibited for the NIH Peer Review Process (NIH, 23 June 2023), checked 5 November 2026 through the Internet Archive copy of 24 October 2026; the live page blocked automated access ↩ a b
17. UKRI modernises grant assessment for the age of AI (UKRI, 10 September 2026), checked 5 November 2026 ↩ a b c
18. Current position of the ERC Scientific Council on AI (ERC, 1 December 2023, last updated 25 March 2026), checked 5 November 2026 ↩
19. The use of AI in grant proposal evaluation: Guidelines for ERC panel members and remote reviewers (ERC, 23 March 2026) and ERC clarifies limits on AI use in grant evaluation (ERC press release, 24 March 2026), checked 5 November 2026 ↩ a b
20. AI in the Review Process (DFG), with DFG form 4.04, Guideline on the Use of Artificial Intelligence in the Review Process (March 2026) and the DFG's AI overview page, checked 5 November 2026 ↩ a b
21. The SNSF's approach to the use of artificial intelligence in funding applications (SNSF, 28 February 2024), checked 5 November 2026 ↩ a b
22. SNSF guidelines for reviewers and referees (SNSF, October 2026), checked 5 November 2026 ↩ a b
23. Policy on Use of Generative Artificial Intelligence in Grant Applications and Peer Review (NHMRC, 28 April 2026) and Statement from the CEOs of ARC and NHMRC (28 April 2026), checked 5 November 2026 through the Internet Archive copies of 16 July 2026 and 16 July 2026; the live site did not respond to automated access ↩ a b
24. DFG Formulates Guidelines for Dealing with Generative Models for Text and Image Creation (DFG, Information for Researchers No. 72, 21 September 2023), checked 5 November 2026 ↩
25. New guidelines for the use of artificial intelligence in peer reviews (SNSF, 20 August 2026), checked 5 November 2026 ↩
26. Carla Carbonell Cortés et al., AI-assisted prescreening of biomedical research proposals: ethical considerations and the pilot case of "la Caixa" Foundation (Data & Policy, 23 October 2024), checked 5 November 2026 ↩ a b
27. Use of AI tools in IGoR proposal reviews (ARPA-H) and ARPA-H launches new program to deliver rigorous, gold-standard research faster (ARPA-H, 5 May 2026), checked 5 November 2026 ↩
28. Frances Jones, AI-only research grant rejection in UKRI-funded call sparks alarm (Times Higher Education, 25 September 2026), checked 5 November 2026 ↩
29. Frances Jones, Grant proposals rejected by 'AI triage' to be reassessed (Times Higher Education, 5 October 2026), checked 5 November 2026 ↩
30. Living guidelines on the responsible use of generative AI in research, fourth version (European Commission, ERA Forum, 15 October 2026), checked 5 November 2026 ↩ a b
31. Diogo Machado, Generative AI bias against scientific novelty: a cautionary tale from a small-sample evaluation of research proposals (Scientometrics 131, 7479–7497, August 2026), checked 5 November 2026 through the abstract the publisher deposited with Crossref; the publisher's page blocked automated access ↩ a b