IN DEPTH ‘‘
Rob Mackinlay is a journalist for Information Professional.
While librarians are wary of many aspects of AI, using it to evaluate research could help undermine the prestige publishing culture.
Has your AI model got your morals?
Where AI morals, models and tools clash with power, control and HE research, discussed with Jisc’s Michael Webb, Director of AI, Lawrie Phipps, Director of External Policy Development, and Rebecca Flook, Principal AI Specialist.
BIG ethical clashes over AI have seen AI companies fall on both sides of the moral divide. Grok’s sexualised images on one side and Anthropic’s refusal to carry out mass surveillance or operate weapons systems on the other.
But whichever side of the divide they are on, the power to maintain or change that position has remained in their hands. Or so it seems. In the “Who controls AI?” session at Jisc’s Digifest conference Michael Webb, Lawrie Phipps and Rebecca Flook from Jisc unpicked the levers of power in AI. “There’s a broad understanding that the training data contributes to bias, I think we understand that,” Michael said. “There’s much less understand- ing about the other levers and technical approaches that the companies like open AI and Google have to control their models.”
He sees three levers of control:
l The model – or the algorithm; l the tool – which connects the algorithm to its user; l neutrality – to give users confidence in the tool.
Model
“Once they’ve trained the basic model on the language stuff,” Michael says, “they fine tune it to provide answers that align with their worldview and their values – or those that they think are market- able and acceptable. These are largely guided by
30 INFORMATION PROFESSIONAL
principles that the groups have set and it’s kind of impossible to inspect this and see.”
The character of these moral fine-tunings can be seen if asked the right questions. Michael said: “I pose a moral quandary – I have three cats, butI can only afford two. Which one should I pick to go?” He said all the models were “empathetic” about his cat problem but they had different approaches when it came to solutions. “The US-based model started giving practical advice about how to choose which ones to get rid of,” he said… “The Indian model had a very different value-set and talked to me about how the community could come together to help me solve the problem.”
The tool
He said that the next layer of control is the ‘tool’ – the interface between a human and an algorithm. He compared ChatGPT and DeepSeek. Both models have similar world views (for example, if asked to give a one-word definition of what underpins democracy they both say ‘trust’) but one operates in China, the other in the west. “We don’t get our responses directly from the model, we use the tool. This is really important because it has multiple moderation layers that operate before and after it gives the response. These make political interference much easier. To see this in action ask DeepSeek about Taiwan. It will answer, then it will reflect on its answer through the moderation and
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
