UNESCO Survey: Artificial Intelligence in Higher Education — Global Challenges and Opportunities
Source: UNESCO — Artificial Intelligence in Higher Education: Global Challenges and Opportunities Published: 2025 Publisher: United Nations Educational, Scientific and Cultural Organisation (UNESCO)
Survey overview and methodology
This report presents findings from a survey of 400 respondents drawn from UNESCO Chairs and UNITWIN Networks across 90 countries, with a 38% response rate and a regionally balanced sample. UNESCO Chairs and UNITWIN Networks are designated academic leadership positions and inter-university cooperation networks — the respondents are senior academics and institutional leaders, not general staff. This is an elite-institution sample, and figures on policy readiness likely overstate the position across higher education systems as a whole.
The survey is empirical and descriptive: it documents what is happening across global HEIs, not what institutions must do. This distinguishes it from the UKRI AI Strategic Framework or the UoS GenAI Principles documents, which are normative and prescriptive. Sheffield’s relevance to these findings is therefore about positioning — understanding where the institution sits relative to global peers — rather than about compliance obligations.
Finding 1: Near-universal adoption for research and writing — teaching adoption at the halfway point
Nine in ten respondents use AI tools in their professional work, primarily for research and writing tasks. Approximately half are experimenting with AI in teaching contexts. The gap between these two figures is significant: research and writing use is already mainstream; teaching use remains in the experimental phase.
This distinction matters for how the catalogue is positioned. The tools assessed here are for research workflows — the already-mature adoption context, where use is happening whether or not institutions have provisioned for it. Teaching use is a separate and still-emerging domain.
Sheffield relevance: The 90% professional adoption figure is consistent with the assumptions underpinning the landscape map overview. It validates the case for moving quickly from awareness to institutional provision. Sheffield’s researchers are already using AI tools with or without institutional sanction — the risk is ungoverned adoption, not under-adoption. The catalogue’s research focus is well-calibrated to the higher-adoption context.
Finding 2: Policy frameworks are developing faster than they have been deployed
Nineteen per cent of institutions surveyed already have a formal AI policy; a further 42% report frameworks under development. Combined, 61% have or are actively developing policy. That leaves 39% globally with neither policy nor plans — a substantial governance gap, even within this elite sample.
The survey measures policy existence, not tool procurement frameworks. Having a principles document is a necessary but insufficient condition for governed AI use; the harder work is translating principles into curated tool guidance and procurement criteria that academics can act on.
Sheffield relevance: Sheffield’s existing GenAI Principles for research and productivity place it firmly in the global 19% minority with deployed formal policy — a strategic strength relative to peers. The gap at Sheffield is not policy but operationalisation: translating those principles into a curated, maintained tool catalogue. This is precisely what the current project delivers. See uos-genai-principles-research for the principles that the catalogue’s governance fields are designed to operationalise.
Finding 3: A global governance gap — regional disparities in AI readiness
Approximately 70% of institutions in Europe and North America have or are developing AI guidance, compared with around 45% in Latin America and the Caribbean. Given that the UNESCO Chairs sample overrepresents well-resourced institutions, these figures almost certainly overstate readiness across those regions’ HE systems as a whole.
The practical implication runs beyond headline statistics. When Sheffield researchers collaborate with international partners, share research data, or co-author publications using AI-assisted workflows, they cannot assume equivalent policies on disclosure, attribution, or bias mitigation apply at partner institutions. This is a research integrity and data-governance risk that is not currently captured in the tool schema.
Sheffield relevance: Sheffield’s research partnerships span the Global South extensively — BRICS collaborators, African Union networks, and LAC partnerships are all operating in contexts where the UNESCO data suggests a 30–55 percentage-point gap in policy readiness relative to European institutions. The existing schema fields data-residency and gdpr-compliant track data flow risks but have no equivalent for partner-institution AI governance exposure. This is a schema gap candidate: the tags field could be used to flag tools as international-partnership-risk where cross-border collaboration with under-governed contexts is likely, without requiring a schema version increment.
Finding 4: Ethical issues are already live — not theoretical
One in four respondents reported that their institutions had already encountered AI-linked ethical issues. The specific types named include student overreliance on AI tools, authorship and attribution disputes in publications, and bias in research outputs. These are reported, recognised incidents at senior academic institutions — the 1-in-4 figure is almost certainly an undercount, as it reflects only incidents that were identified and surfaced.
Each issue type maps to existing schema governance fields:
- Student overreliance maps to
human-oversight-required(which the schema sets as alwaystrue) and tolearning-curve— tools with a very low learning curve may be over-adopted without the critical evaluation their outputs require. - Authorship and attribution disputes map to
output-explainabilityand to UoS GenAI Principle 2, which addresses researcher responsibility for disclosed AI use. - Bias in research outputs maps directly to
bias-risk-notes.
The UNESCO data validates these fields as responses to live operational risks, not precautionary ones.
Sheffield relevance: The UoS GenAI Principles for Research already address attribution (Principle 2) and accuracy (Principle 3) — the internal Sheffield equivalents of the authorship and bias issues the survey identifies. Sheffield’s governance framework is substantively responsive to the risk landscape. However, the UNESCO evidence strengthens the case for treating bias-risk-notes, output-explainability, and human-oversight-required as primary procurement criteria rather than optional annotations. Tools where these fields remain unknown or incomplete should not advance to catalogue-status: adopt without a governance review.
Implications for this project
| UNESCO finding | Catalogue implication |
|---|---|
| 9 in 10 already using AI professionally | Validates urgency; Sheffield’s risk is ungoverned use, not non-adoption |
| 61% have or are developing policy | The catalogue fills the gap between Sheffield’s existing principles and tool-level procurement guidance |
| 39% globally have no policy or plans | Sheffield’s pace is appropriate; the next step the sector has not taken is a maintained tool catalogue |
| ~45% LAC vs ~70% Europe have guidance | Consider international-partnership-risk tag (using existing tags field) for tools used in cross-border collaborations with under-governed partners |
| 1 in 4 institutions report live ethical incidents | bias-risk-notes, output-explainability, human-oversight-required should be mandatory-complete before any adopt ring assignment |
| Authorship disputes named as a specific incident type | Tools with output-explainability: low warrant an explicit caveat note; weight this field more heavily in assess/trial decisions |
| Teaching adoption at ~50% vs research at ~90% | Research catalogue scope is correct; teaching-context AI assessment is an explicit future workstream, not an oversight in the current phase |
Recommended adjustments to the catalogue
Three concrete proposals are suggested by these findings. None require immediate action in this commit — they are documented here as a record for future sprint planning.
1. Governance fields as quality gates
Add a validation rule to _scripts/validate-schema.py: tools should not hold catalogue-status: adopt if bias-risk-notes is empty or unknown and output-explainability is low. This is a validator change, not a schema change, so it does not require updating all existing tool notes.
2. International partnership risk tag
Use the existing tags free-text field to annotate tools as international-partnership-risk where data-sharing with partner institutions in under-governed contexts is a realistic use case. This avoids a schema version bump while capturing a risk the current 27-field schema does not explicitly track.
3. Explicit scope note for teaching use
Add a sentence to research-ai-landscape-map-overview.md (or the root README) confirming that AI in teaching contexts is a future workstream, intentionally out of scope for the current phase. The UNESCO data confirms research use is the right initial focus; the scope note prevents stakeholders from treating the catalogue’s research orientation as a gap rather than a deliberate decision.
Drafted: 2026-03-30 · Branch: 8-add-unesco-survey-on-hei-ai-guidance-to-knowledge-base