Mind the AI policy gap: a policy analysis across global funding bodies

Synthesis: Chan (2026) analyzed 56 AI-related policy documents from 46 research funding bodies across the US, the UK, Europe, Australia, Singapore, Mainland China and Hong Kong, and found that only 12 of the 46 had a policy written for funding applications and peer review. Human-centeredness and safeguarding dominated the corpus while AI Governance and AI readiness appeared only in general institutional policies. The contribution is a map of a policy layer that national AI strategies do not reach — and a record of what funders publish, not of what it changes.

Key Findings

1. The layer is thin. Of 117 funding bodies screened, 46 had any AI-related policy, producing 56 documents. Only 12 of the 46 addressed AI directly in funding application and review, producing 13 funding-specific documents from the US, the UK, Australia and Europe.
2. Asia is absent where it counts. Mainland China and Hong Kong supplied 32 of the 117 bodies but only two had a general AI policy and none had funding-specific guidance. Of the 46 bodies with any policy, 24 were American, 8 Australian, 6 British, 4 Singaporean, 2 European and 2 from Mainland China and Hong Kong.
3. Five themes, unevenly carried. Human-centeredness, safeguarding, embeddedness, governance and AI readiness organized the corpus. Human-centeredness was most prominent (42 codings in funding-specific policies, 70 in general), safeguarding next (32 and 63), embeddedness third (11 and 22). Governance appeared only in general policies (46 codings) and AI readiness only there (26), present in just 16 of the 46 documents.
4. The missing themes are the machinery. No funding-specific document contained a governance code — no oversight structure, monitoring process or data-management provision — and none contained an AI-readiness code covering training, infrastructure or research capacity. The instruments that bind applicants regulate conduct; the instruments that describe governance do not bind them.
5. Disclosure is the operative condition, and it is specified inconsistently. Applicants may generally use AI if they disclose it: the National Endowment for the Humanities requires them to indicate clearly where AI-generated outputs appear, Cancer Research UK requires the model and the prompts, and other funders ask only for a general statement. Reviewers are commonly barred from uploading applications and from using AI to rate proposals; the European Research Council sets "non-delegation" as a principle.
6. The coding is checkable. A second researcher independently reviewed 312 coded extracts, agreeing on 288 and disagreeing on 24, an overall agreement rate of 92.3% with Cohen's kappa of 0.90.

Why the funder layer does not reproduce national AI policy

The sharpest result is a mismatch with the literature. Studies of national policy — Saheb and Saheb's analysis of 57 documents from 24 countries, for instance — find governance and technical themes at the center. Funders put ethics and safety there instead, and Chan explains the difference by purpose and audience: a national strategy governs systems, while a funder governs a transaction. What travels down that chain is not the principle but the condition. A funder can endorse human-centered AI in a general policy without turning it into anything an applicant must satisfy, and most did: 46 with an AI policy, 12 with a funding-specific one.

The institutional pattern predicts where rules appear. Funding-specific policies cluster in dedicated research agencies whose mandate is allocating money — the National Science Foundation, the National Institutes of Health. General policies cluster in multi-mission departments where research funding is one function among several — NASA. An agency whose business is grants writes grant rules; a department that also funds research writes the policy its own operations need.

The regional finding reaches the same absence by another route. Chan reads the empty Asian column as centralization: where AI Regulation in Education arrives as state directives across sectors, funding bodies operate inside them and feel less need to write their own. The consequence for a researcher is concrete: a Hong Kong applicant may face no funding-specific AI condition while a US collaborator faces disclosure requirements, reviewer prohibitions and a contract that carries them. The paper argues this asymmetry could disadvantage Asian institutions in collaborations — an inference from the distribution, not a measured outcome.

What the two missing themes cost

Policy is content; governance is machinery — who decides, who monitors, how a rule is revised. The coding counts make that separation visible across document types. Governance and AI readiness score nothing in the funding-specific column and 46 and 26 in the general one. The rule that reaches the applicant arrives without the apparatus that would make it enforceable or revisable, while the apparatus sits in a document written for the funder's own staff.

A fair objection comes from the paper's own embeddedness theme. Funding-specific policies incorporate existing research-integrity, intellectual-property and data-protection rules by reference — UK Research and Innovation situates its generative AI policy inside its good-research-practice policy, and the National Health and Medical Research Council draws on Australia's Privacy Act 1988. Governance may be delegated rather than absent. What the corpus cannot show is whether delegation works: a reference is a claim about machinery, not evidence of it.

The disclosure variance is the symptom. When no process fixes what an AI-use disclosure must contain, each funder invents a standard — prompt-level documentation at one agency, a sentence at another. The form is the only place where a funder's definition of responsible use becomes operational, so its granularity is the policy. A form asking for a general statement cannot support a judgment about whether the AI did the intellectual work, and a reviewer who suspects undisclosed use is told by UK Research and Innovation and Barts Charity not to speculate about it.

Absent AI readiness cuts differently. Readiness codes cover training, platforms, data infrastructure and research capacity; no funding-specific policy mentions them. If funders expect AI use but never fund the capacity, the cost lands on the applicant's institution, unevenly — a distribution the paper does not measure, though no funder made readiness a condition of anything.

What a policy corpus can and cannot establish

The unit of analysis is the document, not the grant. Fifty-six published policies carry no information about how many applications disclosed AI use, how reviewers treated a disclosure, or whether research integrity changed. Every causal phrase — that AI use in high-impact research "may not be adequately regulated," that divergent policies risk "fragmenting research agendas" — is an inference from the texts.

Absence is the weakest signal here. A policy that exists only inside a grant template, a reviewer briefing or an award contract would not appear in a website crawl of navigation tabs and internal search; the crawl is the method, run in March 2025 and refreshed in May 2026. "No funding-specific policy" therefore means "none published where a researcher would find it" — meaningful for applicants, weaker for governance.

The sample frame limits the regional claims. The 117 bodies came from the research services of eight publicly funded universities in Hong Kong, so the 46 are the funders visible to those intermediaries, not a census. Europe contributes 3 bodies to the initial list and 2 to the policy count, which the author calls too small. The Asian finding rests on 32 bodies from one territory and its mainland neighbor, not on Asia.

What remains is a description. Using document analysis, Chan establishes a distribution — how many funders publish AI rules, of what kind, carrying which themes — with coding reliability at 92.3% agreement and κ = 0.90. That is the right evidence for what exists, and the wrong kind for what works.

What this means for practice

• Write the funding-specific document, not just the institutional one. A general AI policy addresses a funder's own staff and is not a condition of award. Most bodies stop there, leaving applicants to infer expectations from documents not written for them.
• Specify disclosure at the granularity the decision requires. Name the model, the prompts and where in the proposal output appears, as Cancer Research UK does. A general statement gives a reviewer nothing to check and invites the speculation several funders prohibit.
• Keep applicant and reviewer rules separate. Permission with disclosure for applicants and non-delegation for evaluators address different risks: authorship and data privacy on one side, judgment on the other. One rule for both mis-specifies what is protected.
• Fund the readiness line if you require the capability. No funding-specific policy in the corpus mentions training, infrastructure or research capacity. A funder expecting AI-assisted proposals without funding the capacity shifts the cost to institutions and rewards those already equipped.
• For researchers, read the funder's funding-specific policy before the institutional one. Where none exists, the operative constraints are the general AI policy plus existing research-integrity and data-protection rules.

Limitations

The authors' limits are the sample and its frame: six regional groupings, bodies identified through eight Hong Kong universities' research services, and a European subset of 3 that the conclusion calls too small. The data are a snapshot — March 2025, refreshed May 2026 — of documents the paper says should not be treated as fixed regulations. A second researcher independently reviewed the coding, with the lead author retaining final interpretive authority, so the themes are a reading rather than a measurement with an error term. Beyond those, the design cannot establish whether any published policy is applied: no review records, no applicant or reviewer data, no comparison of funded and unfunded proposals. The paper's implications for higher education are its weakest link: it argues that funding priorities shape which educational technologies are built and validated — plausible, and untested by the corpus.

Citation

Chan, C. K. Y. (2026). Mind the AI policy gap: a policy analysis across global funding bodies. International Journal of Educational Technology in Higher Education, 23(60).