ARMA REF SIG drop-in session: AI for REF preparation by ARMA UK

Ethics of using AI in REF

acresswell6 — 2025-09-15 10:16:58 UTC

There are many ethical considerations these are just an handful of thoughts...

Lack of transparency about how decisions were made and conclusions were come to
AI may amplify biases resulting in wrong / discriminatory evaluation outcomes
Privacy & Security - can we trust corporations with our data or would the sector need to build the AI that would be used for REF?
Access and availability of AI tools may vary across institutions (unless a central AI tool is made for all the sector). Will large institutions that can afford staff to input data into AI (or specialists in AI prompts) see large institutions
Can we be certain of its accuracy. Peer review was always seen as the gold standard. Has AI really got to the point where it has overtaken peer review, or is it just a cheap alternative?
Reviewer biases; will reviewers score work generated by AI worse than that fully created through human endeavour?
Environmental impact of AI use (high energy and water consumption in data centres and increased churn of electronic waste as AI hardware is constantly updates). This runs counter to years of work done in the sector to make ourselves more sustainable
What is the endgame? Do we really want human processes automated and replaced by AI?

Research Opportunity

lstennett — 2025-09-25 13:35:09 UTC

I think if we have a secure LLM (meaning codes of practice, outputs, impact case studies, people, culture and environment statements would not go into a public training model, at least if/until they were made publicly available) then I think there's an opportunity to conduct some research. With the right prompt engineering, we could compare the scores and feedback from the panels with what the LLM says and perform comparisons. Humans have all sorts of conscious and unconscious biases and therefore LLMs will too - we would not want to perpetuate or exacerbate these, but a genuine comparison would be interesting as test research without actually forming part of any scoring, official assessment or feedback.

Future-proofing rules?

2025-11-03 11:35:06 UTC

More of a comment about REF rules. We're three years away from submission. I don't see how the REF team will keep up with the inevitable changes in AI capability, AI literacy, tolerance of AI, etc etc.

Define use of AI?

2025-11-03 12:16:01 UTC

AI is a very wide topic. I think we need to be clear what we are talking about. Different tasks have different levels of risk, from full blown output assessment to simple text editing of narrative elements.

Validate peer review scoring patterns (e.g., detect “hawk/dove” pairs, disciplinary scoring differences, or outlier scores).
Combine peer review scores, citation data, altmetrics, and past REF scores to inform output selection.
Identify publications authored by staff, even when names are inconsistent, or co-authored outputs are involved.
Analyse representativeness of submissions using Fields of Research and other indicators.
Flag potential eligibility issues (e.g., outputs not meeting REF criteria).
Identify additional evidence for an impact case study. Extract and link supporting evidence to specific impact claims.
Use NLP to track mentions of research in policy documents, media, patents, or clinical guidelines
Ensure case studies comply with REF guidance: word limits, headings, and structure.
Evaluate draft case studies against REF 2021 examples and suggest improvements. Highlight gaps in evidence or narrative that may weaken the impact case.
Generate draft impact summaries from templates.
Edit and redraft text for clarity, conciseness, and readability (e.g. impact, environment, 100 words).

Developing ICS

2025-11-04 10:34:21 UTC

The pros/cons of using AI to find and create narratives of case study development: we are surveying about 50 impact leads across UCL to form guidance and guard rails

Inequity

2025-11-04 10:40:06 UTC

Are we at risk of creating an uneven playing field for REF where funds dictate whether an HEI has access to safe AI space via subscriptions and others might risk exposing data if they are unable to afford a more secure AI space?

Awareness and Understanding around AI

holliejopen — 2025-11-04 10:56:10 UTC

My main concern is that there are many people beginning to use AI tools without a full awareness or understanding of how these tools work and what happens with inputted data. Creating policy and guidance is a good first step, but how do we ensure our research communities have the appropriate training to be able to follow these. That initial learning is a big time investment for staff who are already stretched for time.

Brianstorming ideas

2025-11-04 11:01:19 UTC

Undetected uses of AI

2025-11-04 11:04:54 UTC

Databases like dimensions, scopus, et al, have AI tools built-in and so users may be using AI in ways that are not captured or detectable