to change the results and other things to happen around the health research?” He said: “The last thing that you want is research results from AI to start giving you results that have been influenced by politics. The power that is held within the governments is the thing that worries me most. I know Anthropic recently stood up and said ‘no’, but I think under pressure a lot of companies would cave.”
Summarising research “When it comes to doing research, doing teaching, working in a library, these things are going to be abstract and remain abstract because institutions can’t easily influence what’s going on at that level,” Lawrie said. But decisions still have to be made and Lawrie is co-author of the recently pub- lished REF AI Report (
https://chet.bristol.ac.uk/ research/ref-ai-project) which looks at the role of AI in the Research Excellence Frame- work (REF).
Billions of pounds of block grants are allocated to universities on the basis of REF and the report estimates that universi- ties will spend around £1bn collectively on the process of selecting the research to sub- mit for evaluation by panels of academics (
https://2029.ref.ac.uk/panels/main-and-sub-panels/). “They’ll select a whole series of outputs to put forward to the panels,” Lawrie said. “These panels, which are made up from the sector, will judge that research against a set of guidelines and decide where it falls on the spectrum of quality.” Surveys carried out for the report found that AI was already widely used by academics: “We heard a lot about how AI was being used to summarise papers or to provide feedback – the rights and wrongs of that, I’m not debating – this is what they’re telling us.”
He said that the potential read across to REF was strong: “When they extrapolate that, they say that because of the volume of research outputs that they will have to evaluate – just internally, to get to the submission process – they were saying ‘why wouldn’t I use AI to read it and to tell me this is better than that, and that’s better than this’?
“And the assumption by some people was ‘why wouldn’t the panels?’ That’s the fear on which we were recommending quite strong guardrails to ensure that if AI is used, that it’s used correctly. But the biggest thing that I heard from academics is that they want transparency about how it’s being used.”
Summary justice
“Even summarisation, which feels like it’s really neutral because it’s not getting information from anywhere else, actually contains biases,” Michael said. “Some areas of this are well researched, and some not. The more researched areas are things like the tendency towards general-
32 INFORMATION PROFESSIONAL US Department of Defense clashes with AI providers over autonomous weapons.
isation. AI will often pick up something that’s only mentioned in a document in passing and amplify it in the summary because it’s not good at understanding the context of what’s important and what’s not. “Far less research has been done on what points it chooses to pick out when it summarises… yet people are treating it as if it was a perfect neutral tool.
“People aren’t aware of this and it means the guardrails that Lawrie described are really important.”
He said people needed to see it to believe it: “If you give an example, people then know what to look at.
“But if you’re saying an abstract, ‘Hey, you shouldn’t do this, because it might be biased’ then people will go: ‘It (the AI summary) looks all right to me. I’ll crack on and use it anyway’.”
Lawrie said: “That’s the issue that we’ve got with the REF. We haven’t got to the stage where people openly say ‘this is what I’ve done, and this is the problem I’ve had with it’.”
He goes on to say: “I would definitely like to see the data on both sides. I’d like to see somebody do a manual judgement and I’d like to see how the AI tools cope with it, which models cope better with it,” adding, “we have to bring people like librarians into those conversations. They know how to evaluate, and they have critical skills.”
Gold standard
Both Lawrie and Michael say human judgement needs to be compared with the tool’s judgement. But not just to see where it fails. “This is beyond the REF,” Michael said. “This is about using AI for all evalu- ative purposes. There’s almost a tendency
to think that humans are really good at that, that they’re the gold standard that we want to hit. That’s probably not the case. We know that human judgement is based on time of day, what they had for lunch, all of those kind of complex things.” Lawrie gives an example: “I’ve got a colleague in a Russell Group university who was given 48 research papers to read that week to judge which ones should go forward to the REF. What’s happening by the time you get to number 40? If it was me – I’ve got dyslexia, so I actually do take a long time to read a paper – by the time I get to number 20 I’m already strug- gling. The system says you must treat all papers the same, but we know that humans get tired and their mood changes, and their behaviours.”
Prestige problem
While librarians are wary of many aspects of AI, using it to evaluate research could help undermine the prestige publishing culture. Lawrie said: “If we ignore AI, what we’re talking about is using tools to judge the quality of research. And those tools have got inherent biases. I’m inherently biased. I will read something, and I’ll go, that must be good because it’s in Nature. I’ll be honest, that would be my first instinct. Do other academics do that? I don’t know. My guess is that we’ve all got those inherent biases.”
Michael said: “It’s more the fact that we need to spend time evaluating what AI can add to it in a really sensible way. It’s a reasonable hypothesis that AI tools would be less biased than humans, but we’ve got no evidence on this particular dataset. It’s something that should be funded to explore.” IP
April-May 2026
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53 |
Page 54 |
Page 55 |
Page 56 |
Page 57 |
Page 58 |
Page 59 |
Page 60
