AI and Fair Research Evaluation in Higher Education: What UK Universities Are Already Doing for REF2029
Generative AI is already in use across UK universities to support Research Excellence Framework (REF) work. Adoption varies widely, trust is mixed, and the sector is calling for national guardrails.
- GenAI tools are being used across multiple universities for REF-related tasks
- Usage differs significantly by institution and purpose
- Staff sentiment skews skeptical; support is strongest for impact case studies
- Clear opportunity exists-if paired with oversight, training, and shared access
Why this matters
The REF shapes how roughly £2 billion in public research funding is allocated each year. REF2021 is estimated to have cost £471 million across the sector, averaging £3 million per institution-costs likely to rise for REF2029.
If GenAI can reduce administrative load and improve consistency, that's meaningful. But without standards, it risks widening gaps between well-resourced institutions and everyone else.
How universities are using GenAI now
- Gathering and organising evidence for impact case studies
- Drafting and refining impact narratives
- Building in-house tools to streamline REF workflows
- Assisting internal reviews and scoring of submissions
The pattern is clear: institutions with stronger in-house expertise and budget are pushing further, faster. Others are experimenting at the edges or waiting for formal guidance.
What staff think
In a survey of nearly 400 academics and professional services staff, the majority opposed AI use across most parts of the REF process, with strong opposition ranging from 54% to 75% depending on the task. The most support-around 23%-was for using AI to help develop impact case studies.
Discipline and role matter. Professional services teams and post-92 institutions are generally more open to AI use than academics in arts, humanities, and social sciences.
Leadership perspectives
Pro Vice-Chancellors interviewed ranged from pragmatic adopters to cautious skeptics. Some see AI use as inevitable and want to lead responsibly: "We need to understand it and experiment a bit." Others worry about trust, limits, and overhype: "We might be in an AI bubble; the limits may become clearer over the next few years."
Even among those wary, there's a shared view that ignoring AI isn't a strategy. The conversation has shifted from "if" to "how."
Risks and equity
Without policy, usage will fragment, and advantages will accrue to institutions with custom tooling. Participants consistently called for national standards, common definitions of acceptable use, and transparent disclosure of where AI helps shape submissions.
Human judgment must stay central. AI can assist with synthesis, structure, and consistency-but it shouldn't be the final arbiter of quality.
Practical steps for HE leaders ahead of REF2029
- Publish a clear institutional policy for GenAI in research assessment, including disclosure rules
- Deliver role-specific training for academics, impact leads, REF managers, and panels
- Protect confidentiality and IP: local deployments, access controls, audit trails
- Pilot before scale: define acceptable use cases, measure time saved and quality effects
- Set up governance: risk registers, bias checks, red-teaming, and human-in-the-loop review
- Avoid inequity: share tools, templates, and guidance across departments and partners
- Document everything: prompts, versions, datasets, and decisions for auditability
What a sector-wide framework should include
- Standardised use cases (e.g., impact evidence synthesis, formatting, quality checks)
- Mandatory disclosure of AI assistance in submissions
- Model and data governance: security, bias testing, and provenance
- Audit logs and reproducible workflows for institutional and national review
- Accreditation or approval routes for tools used in REF processes
- Shared, high-quality AI platforms accessible to all HEIs to reduce inequity
For current policy context and updates on REF, see the official Research Excellence Framework site. Sector-wide oversight will likely involve Research England and partners; watch for guidance and consultation windows.
Global context
Comparable national research assessment systems in Australia and New Zealand have been discontinued or rethought. There's a wider pattern here: old models are straining under new data realities. The UK has a chance to set a balanced, credible standard that others can follow.
For academics and professional services teams
- Use AI as an assistant, not a decision-maker-prioritise clarity, evidence, and accuracy
- Focus AI use where support is strongest: structuring and evidencing impact case studies
- Keep your data safe: no sensitive or identifiable information in public tools
- Record your methods: what tools were used, how outputs were checked, and by whom
- Share wins and pitfalls internally so practice improves across the institution
Skill up, safely
If your institution is building policy, you'll need practical training fast-especially for impact leads, REF managers, and research administrators. Curated learning tracks by job role can shorten the ramp-up.
Explore AI courses by job role for structured upskilling and implementation support.
Bottom line
AI won't fix research assessment on its own, but it can reduce workload, improve consistency, and widen access-if the sector moves with clear rules and shared infrastructure. With strong governance and human oversight, universities can protect fairness and quality while spending less time on process and more on research.
As one sector leader put it, the task now is to combine principle with pragmatism: set the safeguards, involve people across roles and disciplines, and use AI where it actually creates value.
Your membership also unlocks:
