When Algorithms Inherit the Academy: The Silent Erosion of Academic Judgement in British Universities

There is a particular kind of knowledge that does not appear in any handbook. It accumulates slowly — across years of reading application essays, sitting on funding panels, arguing over departmental priorities in draughty seminar rooms. It is the knowledge of what a promising scientist looks like before they have become one. And it is, by most accounts, disappearing.

Across Britain's universities, algorithmic and AI-assisted decision-making systems are being adopted at a pace that has outrun serious institutional debate. Admissions offices deploy predictive scoring tools. Research councils pilot machine-learning models to flag high-impact grant applications. University management boards rely on data dashboards to determine which departments merit investment and which face consolidation. Each individual adoption seems reasonable — a modest efficiency gain, a reduction in administrative burden, a promise of consistency. Taken together, they represent something more troubling: the systematic displacement of human academic judgement from the processes that define what a university is.

The Tacit Dimension of Academic Expertise

The philosopher Michael Polanyi coined the phrase "tacit knowledge" to describe the things we know but cannot fully articulate — the craftsman's feel for the material, the clinician's instinct before the diagnosis. Academic institutions are, in this sense, extraordinarily rich repositories of tacit knowledge. A senior admissions tutor who has reviewed ten thousand applications over three decades carries within her a finely calibrated model of potential that no rubric fully captures. A veteran grant committee member knows, from long experience, which research proposals are genuinely bold and which are merely dressed to look that way.

This expertise is not infallible. It is subject to bias, to fatigue, to the distortions of personal preference. These are legitimate concerns, and they have driven much of the appetite for algorithmic alternatives. But the remedy for flawed human judgement is not the elimination of human judgement. It is its improvement — through training, through structured deliberation, through diverse panels and transparent criteria. What algorithmic systems offer instead is the removal of judgement from the process altogether, replaced by pattern-matching against historical data.

The problem, as several researchers in science policy have noted, is that historical data encodes the past, not the future. A model trained on the characteristics of previously successful grant applications will systematically favour proposals that resemble what has already been funded. In a research environment that urgently needs to support genuinely novel inquiry — the kind that does not yet have a track record — this is not a neutral technical preference. It is a structural conservatism built into the infrastructure of British science.

Admissions and the Quantification of Potential

The consequences are perhaps most visible in undergraduate admissions, where the pressure to process large volumes of applications efficiently has made algorithmic screening tools increasingly attractive. Several Russell Group institutions have piloted or adopted systems that assign predictive scores to applicants based on a combination of predicted grades, school performance data, and socioeconomic indicators. Proponents argue these tools reduce unconscious bias and improve consistency. Critics point out that they also flatten the very qualities — intellectual curiosity, unconventional thinking, resilience in the face of adversity — that experienced tutors have long regarded as the most reliable indicators of academic promise.

There is an uncomfortable irony here. Many of Britain's most celebrated scientists and scholars arrived at their institutions with application profiles that would score poorly on any algorithmic rubric. The student who came late to academic achievement, the applicant whose personal statement described a circuitous intellectual journey, the candidate whose school context made raw grades a misleading proxy — these are precisely the individuals most likely to be filtered out by systems optimised for predictive consistency rather than human complexity.

Grant Allocation and the Homogenisation of Science

The stakes are equally significant in research funding. The UK's major research councils — including UK Research and Innovation and its constituent bodies — have shown growing interest in machine-learning tools that can assist with the initial screening and prioritisation of grant applications. The administrative logic is understandable: the volume of applications received by bodies such as the Engineering and Physical Sciences Research Council or the Medical Research Council has grown substantially, and the burden on peer reviewers is considerable.

Yet the introduction of algorithmic pre-screening raises questions that go well beyond administrative convenience. What criteria does the model optimise for? If it is trained on citation impact, it will favour established research paradigms. If it weights interdisciplinary keywords, it may reward the appearance of novelty without the substance. And critically, who is accountable when a genuinely important research proposal is filtered out before it reaches human eyes?

This accountability gap is one of the most underexamined aspects of algorithmic adoption in British higher education. When a human reviewer rejects a grant application, their reasoning — however imperfect — can in principle be interrogated, appealed, and learned from. When an algorithm does so, the decision is often opaque even to the institution deploying it. The expertise that once resided in experienced reviewers is not preserved within the system; it is replaced by a statistical proxy that mimics its surface outputs without inheriting its depth.

Institutional Memory and the Long View

British universities have survived and flourished over centuries in part because they have been capable of transmitting institutional wisdom across generations. The mechanisms for this transmission are informal as much as formal — the conversations between junior and senior academics, the mentorship relationships, the shared experience of sitting on committees and watching how consequential decisions are made and unmade.

Algorithmic systems interrupt this transmission. When decisions are delegated to automated tools, the occasions for the exercise and development of human judgement diminish. Younger academics and administrators learn to interpret dashboards rather than to deliberate. The tacit expertise that took decades to accumulate is not encoded into the algorithm; it simply ceases to be reproduced.

Some universities are beginning to recognise this dynamic. A number of institutions have introduced what they describe as "human oversight" requirements for AI-assisted decisions, mandating that algorithmic recommendations be reviewed by qualified staff before implementation. These are welcome steps, but they risk becoming procedural fig leaves if the humans conducting the review lack the experience — or the institutional authority — to meaningfully override the system's outputs.

Towards a More Considered Adoption

None of this is an argument against the use of computational tools in university administration. Used appropriately, such systems can genuinely assist with workload management, flag potential inconsistencies, and surface information that human reviewers might overlook. The question is not whether to use these tools, but how — and under whose authority.

What Britain's universities need is a framework for algorithmic adoption that takes institutional knowledge seriously as something worth protecting. That means insisting on transparency about how models are trained and what they optimise for. It means preserving meaningful human deliberation at the points in the process where judgement genuinely matters. And it means resisting the temptation to treat efficiency as a value that overrides all others.

The academy's reputation rests not on the speed of its decisions but on their quality. Algorithms can process applications faster than any committee. They cannot, as yet, recognise a future Nobel laureate in an unconventional personal statement. That capacity — imperfect, irreplaceable, painstakingly acquired — deserves more institutional protection than it is currently receiving.