It would be an understatement to say that artificial intelligence (AI) has dominated political conversations over the past few years. At Policy Connect, we’ve discussed AI as part of our range of All-Party Parliamentary Groups and also through our major research reports. Most recently we released Count on It: Numeracy, AI and Social Mobility with KPMG and National Numeracy, examining how AI intersects with numeracy and what that means for social mobility.
As part of that research we spoke to a range of stakeholders, including Doug Specht, Head of the School of Media and Communication at the University of Westminster. In August, I spoke to Doug about his own path into AI research, what it means for how universities prepare graduates, how AI is reshaping the labour market they enter, and where numeracy fits into a picture so often framed in terms of language and literacy alone.
Coming to AI through information
Doug’s background is in geography, though his work has centred less on any single discipline than on the production of information itself, principally through maps, but more broadly in how knowledge is made, who shapes it, and who it ends up representing. AI, in that sense, is simply the latest infrastructure he has found himself examining, another system that produces and organises information, with its own choices embedded in how it does so. Alongside this more conceptual interest, he also engages with AI in a far more immediate way, as Head of a School navigating how it is reshaping teaching, assessment, and what universities expect of their students day to day.
“My interest is how knowledge is made, how it’s represented, and how it gets circulated. AI is just another knowledge production infrastructure, another way of representing knowledge.”
Not just a question of truth
Much of our conversation turned on what Doug sees as a common misconception about AI, that it functions as a knowledge retrieval system, a faster way of finding the right answer to a question. Much of the public conversation about AI reliability focuses on hallucination, and the tools themselves have increasingly optimised for accuracy in response. Doug argued that this framing, whilst not wrong, misses something important. Any AI system retrieving or generating information is making significant judgements about what to include and what to leave out, regardless of whether the AI happens to be correct. He compared it to a map. Someone using a map might reasonably question whether the specific route it offers is the right one, but will rarely stop to ask whether the map itself, in its choices of what to include and omit, is flawed. He suggested the same applies to AI. Even the simplest administrative task, asking a tool to summarise a document or organise a list, involves the tool making decisions on the user’s behalf, well outside their oversight.
“It’s why maps fascinate me as a knowledge system, because we don’t question them. They’re taken as infallible, but they’re completely fallible. Just like AI, they can never get to a point where they’re 100% accurate, because you always have to choose what you don’t include.”
This connects closely to a finding at the heart of Policy Connect’s report Count on It. AI tools tend to present information as complete and accurate regardless of whether it actually is, and rarely signal to a user when a response might be missing something significant. It is comparatively easy to catch an AI tool getting something wrong. It is much harder to notice everything it has left out. Doug illustrated this with a simple example, asking an AI tool to remove duplicates from a list. Any duplicate it fails to catch will be obvious. But if it removes an entry that was never actually a duplicate or removes both items on a duplicated, that omission is invisible to the user unless they already know the full list by heart.
“Take removing duplicates. You can absolutely see the output where it’s messed up and kept two of the same things in, but you can’t see where it’s accidentally deleted both of them. You can only see the errors in front of you.”
Judgement over production
Doug’s account of AI’s hidden decision-making echoes a central conclusion of Count on It. Our report identifies critical thinking, the capacity to interrogate an output rather than accept it, as the single most important skill for engaging with AI safely and effectively. In communications and media specifically, Doug described this shift as one between production and judgement. AI tools can increasingly take on a role in production, drafting, summarising, generating first attempts, though Doug noted that what they produce is often only middling, rarely as good as a skilled person working the material directly. What is left for the human to do, and what increasingly defines the value they add, is judgement, deciding whether what has been produced is any good, accurate, and says what actually needs to be said.
“Judgement becomes the critical skill, the judgement of whether there’s sufficient evidence to include or not include something, those editorial judgements. This has always been the key tool of media and communication, but this really brings it to the fore.”
Doug was clear, however, that judgement of this kind is not innate. It comes from experience, and often from manual practice, working through a task directly rather than delegating it from the outset. Several stakeholders in our own research made a similar point in relation to numeracy specifically, arguing that working through a calculation manually, even imperfectly, builds a kind of mental scaffolding that makes it possible to judge whether an answer is in the right range, scaffolding that is harder to build once a tool is doing the work from the start. Where that scaffolding never develops, an AI output can look just as plausible whether it is right or wrong.
“What frustrates me is that people tell me they want students to be more critical, but they have to go through the process to become more critical. We can get a lecturer to critique something, because they’re critiquing based on knowledge and experience. But you can’t go straight to the critique, because the fundamental learning hasn’t happened first.”
Preparing students for an AI world
That tension between conceptual and practical understanding runs right through his work as Head of School. The whole sector is grappling with how AI is changing the way students engage with learning and how institutions assess it, questions that go well beyond simply adapting to a new tool. Together with his colleague Gunther Saunders, Doug helped develop the ‘FUTURES’ framework, an attempt to set out the capabilities students need to operate well in a world increasingly shaped by human-AI collaboration.
“In an educational setting, we’re asking, what learning are our students doing? Is AI supporting them to produce new knowledge, or is it taking away opportunities to do that?”
‘FUTURES’ is exactly the kind of framework our own report argues institutions should not have to build alone. Count on It recommends a publicly available critical AI literacy curriculum framework, developed nationally and made freely available, so that providers without the same in-house expertise are not left to develop this thinking from scratch.
The skills gap employers can’t name
Our conversation also turned to the graduate labour market, and the mismatch Doug sees opening up within it. Universities have to balance how they market their courses to prospective students against what will actually prove valuable to the employers those students go on to work for, and employers themselves often struggle to articulate what skills they will need, let alone anticipate how AI might have reshaped their sector by then. At the same time, there is a growing expectation that graduates arrive already operating at a higher level, just as funding for in-work development and adult education has been reduced.
“We get more pressure on us, not just from employers but regulators and government, to have ‘job ready graduates’. It’s taking that responsibility away from government or employers and putting it into universities’ hands.”
This is something we heard directly from employers as part of our own research for Count on It. Businesses recognise that AI is reshaping what skills matter, but consistently told us they are struggling to invest in the training needed to build them, even as expectations of graduates and existing staff continue to rise. It is also a topic we explored earlier this year through an event with the All-Party Parliamentary Group for Skills, Careers and Employment (APGSCE), looking at what meaningful employer engagement in curriculum design actually looks like. You can read our analysis of some of those ideas here.
One theme that came up repeatedly in our own conversations with employers for Count on It, and one Doug agreed with, was the erosion of tasks graduates once used to build broader capability. Administrative work like note-taking and transcription was never valuable purely in itself, doing it helped junior staff develop what might be called proxy skills, workplace awareness, attentiveness to detail, an understanding of how an organisation actually functions day to day. Where AI tools can now complete those tasks entirely, that incidental training ground for early-career staff risks disappearing along with the task itself.
“By doing those administrative tasks, you’re essentially rehearsing, and rehearsing is in part the way learning works.”
Why numeracy
Rounding off our conversation, Doug reflected on a broader tendency in AI policy: a pull toward overstating what the technology can do, and toward short-term fixes rather than addressing the harder, structural issues underneath. For him, taking part in Count on It offered a useful corrective, a chance to look closely at one specific, underexamined piece of that picture, numeracy.
Only 25% of UK adults surveyed for our research recognise numeracy as a prerequisite for effective AI use, compared with 42% who say the same of literacy. Doug’s own view echoed this closely. What is needed is not formal mathematical proficiency, but a working grasp of concepts like probability, the kind of understanding that lets someone judge whether an AI-generated figure looks plausible, or whether a stated likelihood is being mistaken for a certainty. He also pointed to something distinct about maths as a subject, a level of anxiety and avoidance reinforced by a wider culture in which being “bad at maths” is treated as unremarkable, almost something to laugh off, rather than a gap worth addressing.
“You don’t need to have a degree in statistics and predictive modelling to know what’s going on with AI. But you do need to have a concept of things like probability.”
It is precisely this asymmetry that Count on It sets out to address. Numeracy does not currently appear as a named requirement in AI strategy, education policy, or workplace training frameworks, despite sitting underneath the critical thinking that effective AI use depends on. Naming it, and investing accordingly, is the first step our report calls for.
Across our conversation, from how AI produces and organises information, to the judgement graduates will need in an AI-shaped workplace, to the numeracy that judgement ultimately depends on, Doug kept returning to a single underlying point, that understanding what AI is actually doing when it produces an answer matters more than understanding how it technically works, and that the confidence to question it has to be built deliberately. That is also, at its core, the argument Count on It makes. Numeracy will not make AI trustworthy on its own, but without it, the tools with the greatest potential to close gaps in opportunity risk widening them instead.
You can read the full Count on It report here, and find out more by contacting Rhiannon Tuckett-Jones (rhiannon.tuckett-jones@policyconnect.org.uk).
Related content
Research 