UKCORI: Review and Analysis of UK Institutional Guidance for AI Use in Research
Source: UKCORI — Analysis of UK institutional policy and guidance on AI in research Published: April 2026 (evidence base: open-web searches, September 2025) Author: Dr Lesley Uttley Publisher: UK Committee on Research Integrity (UKCORI), commissioned alongside the UKRI Research Integrity Working Subgroup Licence: CC BY-NC 4.0
Scope and methodology
The report analyses publicly accessible AI strategy, policy, and guidance from a sample of 44 UK universities and 23 national bodies (research funders, councils, learned societies, government-facing groups), drawn from across the four nations and a range of disciplines. Searches were limited to open-web content available in September 2025; absence of public guidance was explicitly not treated as evidence that no internal policy exists.
The study is descriptive and pattern-finding rather than evaluative — it does not name individual institutions or rank them. Its purpose is to identify provision patterns, commonalities, and fragmentation, mapped against the UK Concordat to Support Research Integrity and seven UKCORI themes.
This is the most comprehensive sector-level snapshot of UK institutional AI guidance to date. It is the natural sector-wide companion to Sheffield’s own GenAI Principles for Research and Productivity, and complements the UKRI sector-strategy view in ukri-ai-strategic-framework.
Headline finding: provision is uneven and predominantly student-facing
Across the 44 universities sampled:
| Tier of provision | Share | Description |
|---|---|---|
| Student / assessment guidance only | 61% (27/44) | Education-focused; no separate research provision |
| Research-specific guidance | 27% (12/44) | Discrete from education guidance, addresses research integrity |
| Detailed researcher-competency guidance | 14% (6/44) | Goes beyond permission frameworks into informed-user support |
| Fully integrated framework + policy + guidance | 2% (1/44) | One institution only |
Among the 23 national bodies, 78% (18/23) had a strategy, policy, or think-piece on AI in research — covering grant applications and peer review, responsible use, accelerating UK science, sustainability, and AI governance.
The strategic disconnect: national bodies frame AI as an opportunity (capability building, competitiveness, innovation); institutions predominantly frame AI as a compliance and misconduct risk. Where institutional policy exists it often overlaps with academic-integrity frameworks and is inconsistently applied across departments, with some content sitting behind staff-only portals.
Three universal principles
The report identifies three principles that are present across nearly all guidance reviewed:
- Accountability — humans retain ultimate responsibility for integrity, accuracy, and ethical conduct. AI is a tool requiring critical oversight. AI systems cannot be authors, and their use in confidential peer review is prohibited.
- Transparency — substantive AI use must be openly declared, supporting open-science evaluation.
- Maintaining standards of integrity — AI use must comply with existing ethical, legal, and regulatory standards (data protection, IP, confidentiality).
These are the same three pillars that Sheffield’s GenAI Principles encode as Attribution, Accuracy, and Accountability — the convergence is strong evidence that Sheffield’s existing framework sits firmly within emerging UK consensus.
Risks consistently recognised across the sample
| Risk | Description |
|---|---|
| Data security and confidentiality | Public AI tools breaching GDPR / confidentiality when fed sensitive, personal, or unpublished data |
| Intellectual property and copyright | Inadvertent IP transfer from uploads; AI outputs replicating protected content |
| Accuracy and bias | Plausible-sounding fabrications, hallucinated citations, inherited training-data bias |
| Reproducibility | Probabilistic generation produces divergent outputs from identical prompts |
| AI-detection software unreliability | Broad scepticism — flagged as discriminatory, especially against non-native English writing. Many institutions now explicitly prohibit detection tools, favouring human-led review |
Critical gaps in current institutional guidance
The report identifies systemic under-coverage of cross-cutting integrity concerns:
- Cognitive deskilling — maintaining human reasoning, critique, and analytical skills under AI assistance
- Equity of access — to AI tools and to upskilling opportunities, across career stages and roles (technical staff, professional services)
- Ethical considerations of model training — embedded assumptions, biases, generation mechanisms
- Scientific validity, reproducibility, and epistemological diversity — particularly in AI-enabled research
- Environmental and socio-economic consequences — energy/water consumption, data-centre material costs, labour conditions in data-labelling supply chains, “data colonialism” and Global South impacts
- Career-pathway impact — particularly on entry-level research positions and ECRs
- Non-STEM disciplinary needs — guidance is overwhelmingly STEM-framed; arts, humanities, qualitative social sciences are under-served
- Inclusivity — differential impacts on researchers with different languages, backgrounds, working patterns, accessibility needs
Discipline-specific concerns flagged
| Disciplinary cluster | Specific concerns identified |
|---|---|
| Arts, design, humanities | IP, authorship, creative-field job displacement, aesthetic homogenisation |
| Law, politics, social sciences | Legal implications (IP, criminal law), democratic-process impacts, AI as both tool and object of study |
| STEM and computer science | Reproducibility, explainability, technical integration |
| Engineering and environmental sciences | Resource cost, sustainability, ecological impact of AI infrastructure |
| Qualitative social sciences | Confidentiality breaches in transcription, oversimplification of cultural and contextual nuance |
The report explicitly notes opportunities to develop guidance for qualitative research, archival work, and textual analysis — areas where current institutional guidance is thinnest.
Enabling practices observed in the most-developed institutions
The minority of institutions with detailed research-specific guidance share a pattern: clear permitted/prohibited use guidance, transparent disclosure requirements, and training plus technical support for safe innovative research. Several have invested in operational infrastructure:
- Centrally managed enterprise AI tools with no data retention for external model training
- Bespoke proprietary secure platforms for working with confidential information
- Approved AI tool catalogues linked to institutional data classification schedules — explicit guidance on appropriate AI use for different data sensitivity levels
- Explicit prohibition of AI-detection tools, with human-led review and discussion as the alternative
The “approved AI tool catalogues linked to data classification” pattern is the closest sector analogue to the catalogue this project feeds. The report names this as best practice; Sheffield is in a position to be one of the few UK institutions implementing it.
The seven UKCORI themes
The report structures findings around seven themes mapped to Concordat principles. The themes most relevant to the catalogue:
| UKCORI theme | Key institutional gaps named |
|---|---|
| Sector governance | Inconsistent alignment of policy with the National Security and Investment Act, EU AI Act, and the (forthcoming) UK AI Regulation Bill |
| Roles and responsibilities | Inconsistent acknowledgement requirements; faculty-level variation; need for human oversight when using AI tools |
| Skills and training | Mandatory training rare; AI literacy and AI critical-thinking under-resourced; horizon-scanning capability is patchy |
| Public understanding | Civic education and societal dialogue under-developed |
| Attribution and ownership | No consensus on AI citation format; IP / proprietary ownership of data and models unclear; recognition for “hidden” labour (curation, prompt engineering) absent |
| Reliability and quality | FAIR data principles, AI tool monitoring, AI output provenance — inconsistently applied |
| Research-on-research integrity | Longitudinal monitoring of AI’s impact on productivity pressure, emerging questionable research practices, and meta-research is largely absent |
Implications for this project
| UKCORI finding | Catalogue implication |
|---|---|
| 61% of universities have only education-focused guidance | The Sheffield catalogue’s research-first framing is correctly differentiated from teaching/assessment guidance; this is a sector gap the catalogue helps fill |
| Three universal principles (accountability, transparency, integrity) | Sheffield’s existing GenAI Principles map cleanly onto sector consensus — no schema rework needed; the catalogue should explicitly cite this convergence |
| ”Approved AI tool catalogues linked to data classification schedules” identified as best practice | Validates the core deliverable of this project; consider making the data-sensitivity mapping explicit in the catalogue’s surface UI rather than buried in data-residency metadata |
| AI-detection tools widely prohibited | Out of scope for this catalogue; if any detection tool is encountered, default catalogue-status: hold |
| Cognitive deskilling under-covered | Schema candidate: a deskilling-risk-notes or cognitive-load-impact field, particularly for tools that automate reasoning steps (literature review synthesisers, code-completion agents) |
| Equity of access under-covered | Reinforces case for institution-wide licences over departmental pilots (consistent with ukri-ai-strategic-framework and unesco-hei-ai-survey-2025) |
| Environmental / socio-economic costs under-covered | Schema candidate: environmental-impact-notes field, or extend bias-risk-notes to include supply-chain considerations |
| Non-STEM and qualitative-research needs under-served | Capability-area gap: tools for qualitative research, archival work, and creative practice deserve a dedicated capability-area note rather than being scattered |
| No consensus on AI citation format | Catalogue should not over-specify; attribution-required should remain a flag, with notes pointing to Sheffield’s own Attribution principle (uos-genai-principles-research) and any funder-specific overlay |
| ”Hidden” labour (prompt engineering, curation, validation) unacknowledged | Sheffield’s catalogue process itself is an instance of this work; consider documenting the assessment effort per tool to surface it |
| Research-on-research integrity gap | Suggests value in periodic re-review of catalogued tools, not just one-shot assessment — argues for a sustained sprint cadence rather than a closed project |
Suggested adjustments — for future sprint planning
- Cross-link the three Sheffield principles to the three UKCORI universal principles in uos-genai-principles-research — to evidence sector alignment when the catalogue is presented to stakeholders.
- Consider two new schema fields (would require a minor schema-version bump):
deskilling-risk-notesandenvironmental-impact-notes. Both map to under-covered concerns the report explicitly flags as gaps. - Add a non-STEM / qualitative-research capability area to
capability-areas/, addressing transcription, archival work, qualitative coding, and creative practice — currently the thinnest part of the sector’s guidance. - Document Sheffield’s data-classification-to-tool mapping explicitly in the catalogue UI when published — UKCORI names this as best practice and very few UK institutions have done it.
- Adopt UKCORI’s “AI is not an author and may not be used in confidential peer review” position as a default note on tools where peer-review or editorial use is plausible (e.g., literature-review synthesisers, summarisation tools).
None of these require action in the current sprint — they are recorded here for sprint planning.
Related notes
- cmi-strategy-2026 — Sheffield’s Centre for Machine Intelligence strategy; its internal “AI Sandbox” is the closest Sheffield analogue to the “approved AI tool catalogue” best practice this report names
- uos-genai-principles-research — Sheffield’s research-specific principles (the institutional counterpart)
- uos-genai-principles-productivity — Sheffield’s productivity-focused principles
- ukri-ai-strategic-framework — UKRI sector strategy (the national-body view this report contrasts with institutional provision)
- unesco-hei-ai-survey-2025 — global comparator for institutional AI policy readiness
- research-ai-landscape-map-overview — landscape map this catalogue feeds
Suggested citation
Uttley, L. (2025). Review and Analysis of UK Institutional Guidance for Artificial Intelligence in Research. Report commissioned by the UK Committee on Research Integrity and the UKRI Research Integrity Working Subgroup. October 2025.