Institutional AI capability: connecting education, research and governance

How can universities connect AI use in teaching and research with institutional responsibility and sustained support? ERF’s six-dimensional comparison examines ten universities and distinguishes policies, services, practice and outcomes.

What a comparison of ten universities reveals about the organisational conditions for responsible AI use

Educational Research Foundation (ERF) · Comparative research analysis

Publication Version v1.0 · Published 5 October 2026

Original research cut-off: 23 September 2026 · Publication verification: 5 October 2026

AI and Higher Education

This article develops ERF’s English Research Comparison v1.0 into a publication manuscript. Material claims retained below were rechecked against official sources on 5 October 2026. This was a targeted publication review, not a comprehensive new survey of developments since September. Sources are listed below; the editorial verification record is retained in the publication master.

Executive summary

A university’s capacity to use artificial intelligence responsibly depends on the connections between educational purposes, research practice, professional judgement, shared services and institutional accountability. The existence of an AI policy, a research centre or a campus licence establishes only part of that capacity. The harder question is whether these arrangements help staff and students use AI appropriately, evaluate what happens and improve their work.

This article extends ERF’s comparison of five UK universities—Oxford, UCL, King’s College London, Manchester and Edinburgh—to include Arizona State University, the University of Michigan, ETH Zurich, the National University of Singapore and the University of Tokyo. It retains Assessment, Curriculum and Teaching Practice as its educational dimensions and adds Governance, Research and Institutional Capability. The last term is used narrowly for the people, resources and services that sustain implementation.

The evidence reveals different institutional arrangements. Oxford explicitly separates practical support from policy authority. Michigan combines shared AI services with researcher development. ASU supports projects through common infrastructure and requires accounts of their outcomes. NUS links a common undergraduate foundation with rules governing AI provision by instructors. UCL’s education programme brings governance, curriculum, assessment, staff development and evaluation into a common structure. These are arrangements to examine, rather than positions in a league table. [O2; US1,US2,US9,US10; SG1,SG2; U6]

Public sources establish policies, services, selected practices and course designs more readily than they establish sustained outcomes across a university. The comparison therefore offers an analytical framework and specific institutional lessons, while leaving unverified coverage and effects explicit. For Japanese higher education, its central implication is to connect institutional responsibilities and support with disciplinary purposes and evidence from practice.

1. From educational AI use to institutional capability

The original UK comparison asked how generative AI was being incorporated into assessment, curricula and teaching. Those questions remain fundamental. A university must be able to explain what students should learn, what assistance they may use and how their own understanding will be demonstrated.

Extending the comparison reveals additional questions. Who decides which uses are acceptable? Can disciplinary researchers obtain methodological support? Who maintains a teaching application after its initial developer moves on? How are errors, costs and user experiences used to revise provision?

ERF uses Institutional AI Capability to mean the capacity to establish educational and research purposes, combine relevant expertise and resources, evaluate the appropriateness and outcomes of AI use, and sustain or revise that use institutionally. This is a working definition for comparative analysis, not an international accreditation standard.

The framework encompasses AI broadly where institutional research and governance sources do so, while the educational evidence is predominantly about generative AI. Claims concerning AI research organisations are therefore not automatically evidence of generative AI integration in teaching.

2. Six dimensions and a common evidential standard

Table 1. The same six questions guide all institutional cases.

DimensionCentral questionEvidence required for a stronger judgement
AssessmentCan the university specify appropriate AI assistance and assess students’ own understanding?Assignment conditions, marking criteria, process records and oversight of AI-assisted marking
CurriculumIs AI related to intended learning outcomes and disciplinary progression?Curriculum maps, approved course outcomes, progression and evidence of attainment
Teaching PracticeWhat do instructors and students actually do, and how is practice improved?Teaching records, documented activities, feedback and evaluation
GovernanceWho authorises use, remains accountable and can require correction or suspension?Policies, delegated authority, approval procedures and review mechanisms
ResearchCan researchers use AI appropriately within disciplinary workflows?Methodological support, documentation, validation, reproducibility and researcher development
Institutional CapabilityWhat enables implementation to continue?Shared services, specialist staff, recurrent resources, support, evaluation and migration arrangements

Governance concerns decisions and responsibilities. The sixth dimension, organisational enabling capacity, concerns the resources and operations that make those decisions effective. This distinction avoids using the overarching concept as its own component.

Research requires a further distinction between research on AI, research with AI, and services that enable researchers to use AI. A prominent AI laboratory may be relevant to all three, but its existence alone does not demonstrate that researchers throughout the institution receive appropriate support.

The ten cases were selected purposively to examine different arrangements. They do not represent national averages, and most are substantial institutions with resources that many other universities cannot match. The study prioritises official policies, course catalogues, service descriptions, original research and institutional reports. It does not sample every syllabus or inspect internal budgets, contracts, submissions or usage logs.

For each claim, ERF distinguishes scope, date and state: policy or design; pilot; reported implementation; continuing service; and evaluated outcomes. These are descriptors, not a maturity score. A course catalogue establishes intended teaching and assessment. An institutional news story can establish a reported activity without independently proving its effectiveness. Missing public evidence means that a claim remains unconfirmed.

3. What the educational evidence establishes

Table 2. Assessment, Curriculum and Teaching Practice. Course examples establish evidence at the indicated scope; they do not measure the proportion of a university’s teaching covered.

UniversityAssessmentCurriculumTeaching Practice
OxfordAI use in assessed work requires advance written authorisation for that assessment. [O1]The selected example concerns a Foundation Year; wider curricular coverage is unverified. [O5]Conversational revision and independent study were reported in Astrophoria Physical Sciences in 2024–25. [O5]
UCLThree optional categories support clear task conditions; they are due for review in 2026–27. [U1]The AI education programme includes curriculum design and development. [U6]A published repository account describes students critically evaluating AI-generated scientific output. [U2]
King’s College LondonPAIR structures problem formulation, tool selection, interaction and reflection. [K1]It offers an approach for incorporating AI into existing coursework. [K1]Guidance reports piloting in more than 19 modules in 2023–24; current continuation is unverified. [K1]
ManchesterThe policy requires four categories for all summative assessments, effective 1 September 2026. [M1]Initial implementation classifies existing assessments; it does not require wholesale changes to learning outcomes. [M1]Implementation across all assessment briefs has not been independently established. [M1]
EdinburghBIME10082 assesses the AI process within a poster assignment. [E1]Appropriate generative AI use is a course learning outcome. [E1]The 2026–27 specification combines ELM with students’ own investigations; completed delivery is unverified. [E1]
Arizona State UniversityA philosophy-of-science example reports assignments using AI dialogue. [US8]The evidence concerns selected disciplinary courses. [US8]A 2024 report describes philosophical dialogue, simulated patients and optional German conversation practice. [US8]
MichiganInstructors set conditions and communicate them through syllabi and assignments. [US4]A compulsory foundation for all students is not established by the selected evidence.An instructor describes developing Maizey study support using course materials. [US5]
ETH ZurichGuidance places responsibility and disclosure on users and calls for explicit disciplinary conditions. [CH1]Instructors are encouraged to review course objectives in relation to AI. [CH1]Ethel Classic offers course-material chat and handwritten-work feedback; the next generation is a pilot. [CH3]
NUSUnsupervised assessment generally permits acknowledged AI use; prohibition requires suitable supervised components. [SG1]THE1008 is compulsory for all first-year undergraduates from 2026–27. [SG2]The university reports pilot and workshop activity; completion and attainment across the intake are unverified. [SG2]
TokyoInstructors determine task conditions. Course 51292 includes an inquiry journal among assessed outputs. [JP1,JP2]The course links AI use to analysis of the inquiry process. [JP2]The 2026 autumn syllabus specifies dialogue, reflection and process analysis; delivery outcomes remain unverified. [JP2]

Two distinctions emerge. First, common assessment language and curriculum reform are different interventions. Manchester’s policy establishes a university-wide requirement for communicating conditions while retaining disciplinary decisions about the category. UCL’s categories remain guidance, and its broader education programme addresses curriculum and organisational development. [M1; U1,U6]

Second, the object of assessment can include students’ decisions about AI. King’s PAIR guidance, Edinburgh’s poster specification and Tokyo’s inquiry journal each make process visible in different ways. Such designs create opportunities to assess judgement; they do not by themselves establish that students have developed it. [K1; E1; JP2]

NUS provides a different scale of curricular commitment: a common foundation for incoming undergraduates. Its announcement describes THE1008 as a starting point for subsequent development through common and disciplinary curricula. Evidence of the requirement is stronger than evidence of whole-cohort attainment or completed progression through every degree. [SG2]

4. Governance, research and the resources for implementation

Table 3. Governance, Research and Institutional Capability. The entries identify documented arrangements; their institutional effectiveness is a separate question.

UniversityGovernanceResearchInstitutional Capability
OxfordThe support centre distinguishes its role from the AI Governance Group’s authority. [O2]Research policy addresses substantive use, accountability, disclosure and reproducibility. [O4]The AI Competency Centre provides tools, training, advice and research support. [O2]
UCLIts education approach and programme recommendations received Education Committee approval. [U6]ARC supplies digital research infrastructure, expertise and training. [U3]The education programme links organisational workstreams; ARC supplies a further research-support function. [U3,U6]
King’s College LondonThe AI Institute and Research Ethics Office jointly organised researcher workshops. [K3]Doing Science Well with AI supports reflection on responsible research practice. [K3]Shared teaching guidance and researcher development are documented; their resource integration is unverified. [K1,K3]
ManchesterPublished strategy groups include university, research and ethical responsibilities; the teaching group is marked “On-hold”. [M2]The research group’s remit includes infrastructure, collaboration and skills. [M2]The published structure includes AI training and support responsibilities. [M2]
EdinburghInformatics guidance requires ELM-team review before requesting another AI service; central authority is not comprehensively established here. [E2]General research-use support is not established by the sources selected for this article.ELM has dedicated support; AskEdHelp provides an example of service testing and revision. [E2,E3]
Arizona State UniversityCreateAI is described as a governed environment with evaluation and monitoring functions. [US10]The 2024 Challenge report includes activity across research and other settings. [US8]Approved projects receive platform access; successful applicants must report use cases and outcomes. [US9]
MichiganCourse judgement rests with instructors; common services refer users to institutional data controls. [US1,US4]MIDAS provides research resources, tutorials and incubation activities. [US2]U-M GPT, Maizey and GPT Toolkit address different user needs. [US1]
ETH ZurichEducational R&D and central IT’s production service are explicitly distinguished. [CH3]The AI Center has an interdisciplinary research mission. [CH4]Central IT operates a Moodle-integrated platform; Ethel remains a distinct R&D environment. [CH3]
NUSAI instruction, feedback and marking require prior approval through specified academic authority. [SG1]Four AI-for-Science projects connect AI and disciplinary expertise. [SG4]A common educational foundation and institutionally supported baseline tools are documented. [SG1,SG2]
TokyoTeaching guidance assigns decisions to instructors; public service guidance points to internal security requirements. [JP1,JP3]BOOST NAIS supports selected doctoral researchers developing next-generation AI. [JP5]Official guidance describes shared tools as of March 2026; later contractual status is unverified. [JP3]

These arrangements should be read at their stated scope. Edinburgh’s Informatics procurement guidance is not evidence of every institutional approval route. Tokyo’s doctoral programme concerns selected participants rather than support for all research activity. Manchester’s “On-hold” label applies to the named strategy group and does not show that teaching or assessment work has stopped. [E2; JP5; M2]

Shared services and professional judgement

Michigan illustrates how common provision can accommodate different levels of technical expertise. Its services range from conversational use to applications grounded in users’ own datasets and a developer toolkit. MIDAS adds research-oriented resources and skills development. These are complementary functions, although the evidence does not measure uptake or outcomes across the university. [US1,US2]

The practical work of implementation remains visible in the instructor account of Maizey: selecting and preparing materials, testing responses and dealing with errors require attention. That account is evidence of an instructor’s experience, not a statement about the current limitations of every version of the service. [US5]

Oxford makes another useful distinction explicit. The AI Competency Centre supports informed use but does not set university policy. Its research policy, version 1.1 dated 7 September 2026, also identifies review and amendment arrangements. Together, these documents show how support and accountability can be specified without assigning both functions to the same unit. [O2,O4]

Institutional responsibility extends to AI provision by staff

NUS requires approval for deploying AI to deliver instruction, feedback or marks, and leaves ultimate responsibility with instructors. The policy therefore addresses the institution’s provision of education as well as student conduct. Its approval provisions should be distinguished from evidence about the number or quality of systems approved in practice. [SG1]

UCL’s programme connects academic, technical and organisational workstreams, including student participation and pedagogic evaluation. The published structure establishes coordination arrangements; achievement of their objectives remains a separate question. [U6]

Experimentation needs a route to review and continuing operation

ASU’s AI Innovation Challenge requires successful applicants to document their use case and outcomes. CreateAI’s technical description separately identifies evaluation and monitoring functions. These arrangements make review part of the stated implementation process; they do not independently establish the reliability of the evaluations or improvements in learning. [US9,US10]

ETH’s Ethel page distinguishes an established environment, a newer pilot and a central IT platform described as operating in production. That distinction matters when assessing sustainability: an experimental application and a maintained institutional service have different responsibilities. The page does not establish universal use of either. [CH3]

Edinburgh’s AskEdHelp report describes user testing and the use of query records to inform revisions. Because this is a student enquiry service, its findings concern service interaction rather than attainment within a course. It nevertheless offers a concrete example of learning from use. [E3]

5. ERF analysis: capability lies in the connections

The following propositions are ERF’s interpretation of the evidence. They are not the universities’ own comparative judgements.

5.1 Institutional capability is relational

The six dimensions are most useful when they expose connections and gaps. A university may have competent technical provision but limited support for assessment design. It may have carefully expressed policies but no clear process for maintaining applications. Research excellence in AI may coexist with uneven access to methodological assistance in other disciplines. These are diagnostic possibilities, not findings of deficiency at the institutions examined.

Table 4. Connections that a subsequent institutional review should investigate.

ConnectionTest of organisational capacity
Governance → practiceCan staff identify who approves a use, who supports it and who can suspend it?
Curriculum ↔ AssessmentDo the capabilities students are expected to develop correspond to what assignments actually test?
Shared services → disciplinary workCan users obtain technical and methodological help appropriate to their field?
Experience → evaluationAre errors, user concerns and unsuccessful applications recorded and examined?
Evaluation → resourcesDo findings affect continuation, redesign, staffing and future expenditure?

The cases supply examples relevant to these questions, but the full chain has not been audited at any institution. A reporting requirement, for example, is evidence of a review mechanism; proof that reports change resource decisions requires further investigation.

5.2 Common conditions should enable disciplinary judgement

Institutional coordination has particular value for access, data protection, procurement, support and responsibility. Decisions about appropriate methods, acceptable assistance and disciplinary standards require academic expertise. A useful comparison therefore examines how common arrangements support those judgements and how disagreements are resolved.

Uniformity alone is an inadequate objective. A language exercise, a clinical simulation, a mathematical examination and an archival research project have different purposes. The framework should test whether a university can explain and support these differences consistently.

For a student, the work of reaching an answer can be central to developing capability. For a researcher, generating valid new knowledge is a central objective. These purposes can justify different conditions of AI use. Restrictions intended to preserve independent learning can coexist with support for research use subject to appropriate validation and documentation.

Doctoral education occupies both domains. It must assess the validity of the research and the researcher’s capacity to exercise independent judgement. Consequently, research productivity and researcher development should not be treated as interchangeable measures.

5.4 The ability to revise or stop is part of capability

Access, training and project activity are enabling conditions. Stronger claims require evidence about learning, error detection, research robustness, staff workload, cost and equity. Efficiency should include the effort required to check and correct outputs.

An institution also needs to recognise when an application is unsuitable, when a conventional method is preferable and when a service should be redesigned or withdrawn. Treating increased AI use as an automatic success would make that judgement harder to exercise.

6. Implications for Japanese higher education

The comparison suggests a practical agenda for Japanese universities, with adaptation to institutional mission, resources, governance and disciplinary composition.

First, specify responsibility alongside access. A shared service should have a named institutional owner, a support route and conditions for sensitive or consequential use. Academic decisions about learning and research should be connected to those arrangements.

Second, connect common foundations with disciplinary progression. Introductory AI literacy can establish a starting point, but programme teams need to identify what students should subsequently be able to judge, produce and explain. An assignment that records the inquiry process may be useful where its purpose is educationally justified.

Third, support a manageable set of applications and evaluate them properly. Universities could select priority courses or research teams, agree what success and failure would mean, and provide both academic and technical support. Measures should include verification work and user experience, rather than just initial output speed.

Fourth, make evaluation consequential. Findings should inform whether an application continues, how it changes and what resources it receives. Where appropriate, collaborative procurement or shared specialist services could help institutions with limited resources. These are ERF proposals, not universal procedures validated by this sample.

For a subsequent comparative study, common denominators will be essential. The share of assessments with explicit conditions, the proportion of students reaching defined learning outcomes, recurring support expenditure and waiting times for assistance are more interpretable than isolated totals. Accounts, active users, training attendances and individual completers should remain separate measures.

Tokyo’s inclusion provides a specific Japanese point of comparison. It cannot establish the position of Japanese higher education as a whole. Likewise, the five UK cases do not establish a national model against which other systems should be ranked.

7. Conclusion

Institutional AI capability is best examined as a university’s ability to connect purpose, judgement, support, accountability and improvement. The ten cases offer different ways of organising those connections. They also demonstrate why policy, access, course design, reported use and demonstrated outcomes require separate evidence.

The next research step is to examine how these arrangements work over time: who uses the services, how disciplinary decisions are supported, what changes in learning and research can be established, and how evaluation affects institutional choices. That is the evidence needed to move from comparing visible arrangements to judging sustained capability.

Sources

All sources listed below were consulted again on 5 October 2026. Dates refer to publication, update or the period explicitly covered by the source, as indicated. Undated pages are not assigned inferred publication dates. Reference labels retain their relationship to the source research report; gaps in the numbering reflect selection for this article. Japanese titles are translated descriptively.

Related reading

Comparative analysis and editorial responsibility: Educational Research Foundation. Prepared with AI assistance in source checking, synthesis and English drafting; final editorial judgement remains with the ERF Editor.