New AI Framework Puts Students, Not Algorithms, at the Heart of University Admissions

Every year, millions of applicants entrust their futures to university admissions processes that are increasingly being reshaped by artificial intelligence. Machine learning models now help institutions forecast enrolment, flag at-risk students, and even assist graduate admissions reviewers. But a new conceptual study argues that the way AI is being deployed in admissions is fundamentally misaligned with what education is supposed to do. In a paper published in Discover Education, Abdullah F. Alenezi of the Public Authority for Applied Education and Training in Kuwait proposes the Human-Centred Personalised Admissions Framework, or HCPAF, a governance-oriented model that repositions admissions not as a terminal selection event but as the opening move of a personalised learning journey that stretches from application through to graduation and beyond.

The timing of the proposal is significant. Higher education systems worldwide are straining under rising application volumes, constrained resources, and intensifying expectations around widening participation. In the United Kingdom alone, the 2023 UCAS cycle involved more than 750,000 applicants and nearly three million applications. Admissions systems designed for smaller applicant pools and standardised qualifications are buckling under the load, and AI promises scalability. Yet Alenezi’s review of the literature reveals a troubling pattern: most AI-enabled admissions scholarship evaluates systems purely on prediction accuracy, enrolment optimisation, or processing efficiency, while paying scant attention to learner agency, transparency, or educational equity.

The risks are not hypothetical. The paper documents how algorithmic systems trained on historical data can encode and perpetuate existing inequalities. Amazon’s experimental CV-screening tool systematically downgraded applications from women because it had learned from historically male-dominated hiring patterns. The COMPAS recidivism tool demonstrated racially disparate predictions. And the UK’s 2020 A-level algorithm depressed grades for students from lower-performing schools, sparking a national controversy. In admissions, the danger takes specific technical forms: representation bias when training datasets under-represent particular groups, measurement bias when proxy variables like postcode or school type capture group membership rather than merit, and feedback bias when predictions shape institutional actions that then create self-fulfilling cycles.

HCPAF responds with an architecture of five interconnected dimensions: learner profiling and contextual understanding; explainable AI decision support; equity and inclusion auditing; human oversight and shared decision-making; and personalised transition and success support. Crucially, the framework treats these dimensions as recursive rather than sequential. Profiling choices shape what can be explained and audited; fairness findings may force revisions to model features; human review can trigger model correction; and transition outcomes feed back into assumptions made at the point of admission. The result is a system of mutual constraints rather than a frictionless pipeline.

Technically, the framework leans on established explainable AI techniques such as LIME and SHAP, which generate interpretable accounts of how input variables contribute to model outputs. But Alenezi, drawing on the work of Khosravi and colleagues, insists that explainability must be calibrated to its audience. An explanation intelligible to a data scientist may be opaque to an admissions officer or an applicant, so the framework calls for layered explanation architectures serving multiple stakeholder groups simultaneously. All AI-generated recommendations must be accompanied by plain-language explanations, consistent with GDPR Article 22 rights and the EU AI Act’s requirements for high-risk systems in education.

The framework is equally candid about trade-offs that much of the AI-in-education literature glosses over. Richer contextual data can support fairer decisions, but collecting more socioeconomic, demographic, and behavioural information increases privacy risk. Postcode-level measures such as POLAR4 quintiles can identify structural disadvantage, yet the same variables may act as proxies for ethnicity, class, or disability — and removing them entirely can make disadvantage statistically invisible. HCPAF’s answer is neither uncritical inclusion nor blanket exclusion, but counterfactual and subgroup sensitivity testing, with contextual variables used preferentially to expand opportunity rather than deny access. Similarly, the framework acknowledges that demographic parity, equalised odds, and individual fairness cannot always be satisfied simultaneously, and that carrying admissions data into student-support systems may enable earlier help while creating risks of labelling and surveillance.

One of the study’s most distinctive conceptual contributions is its treatment of trust. Rather than assuming that transparency automatically builds confidence in AI systems, Alenezi frames trust as a relational, calibrated, and empirically contingent construct. More information does not necessarily generate more trust: disclosure of uncertainty, data limitations, or unequal error rates may appropriately reduce confidence. The relevant objective is calibrated trust, in which applicants and staff understand what the AI can and cannot do, can challenge its outputs, and have credible evidence that errors and inequities will be detected and remedied. Explainability, fairness auditing, and human oversight may reinforce one another — or they may reveal justified grounds for distrust.

Generative AI occupies a carefully bounded role within the framework. While predictive machine learning supports profiling and fairness monitoring, large language models are positioned mainly as an interaction and explanation layer: translating model outputs into applicant-accessible language, providing source-grounded conversational guidance about programmes and procedures, and personalising pre-arrival information. Recognising that LLMs can hallucinate, reproduce bias, and generate confident but incorrect advice, HCPAF requires retrieval from authoritative institutional sources, logging and audit of outputs, clear disclosure that applicants are interacting with AI, and escalation to human advisers wherever financial, disability, visa, safeguarding, or admissions-condition issues arise. Notably, the framework treats AI-assisted personal statements as a governance problem rather than an automatic integrity violation, urging institutions to avoid unreliable AI-detection tools and to ensure applicants are not disadvantaged by unequal access to language models.

The theoretical scaffolding is deliberately broad, integrating six traditions: Shneiderman’s human-centred AI, responsible AI and algorithmic governance, Sen’s Capability Approach, educational equity theory, Value-Sensitive Design, and human–AI collaboration theory. The Capability Approach in particular reframes the evaluation question: admissions systems should be judged not only by whether they predict academic success accurately, but by whether they expand or constrain the substantive freedoms of diverse learner populations to access and flourish within higher education. The framework also identifies normative alignments with Sustainable Development Goals 4, 10, and 16, though Alenezi is careful to stress that these are alignments of principle, not evidence of impact.

The author is refreshingly explicit about the framework’s limits. HCPAF is a conceptual synthesis, not an empirically validated admissions model, and prospective evidence on live AI-assisted admissions remains thin. The paper’s illustrative scenarios — a post-92 university using contextualised offer-making with monthly equity audits, a research-intensive university deploying a generative AI advising portal with human review of consequential guidance, and a specialist provider connecting admissions profiles to early-warning retention systems — are explicitly hypothetical design propositions rather than documented deployments. Until validation evidence accumulates through case studies, participatory co-design, and longitudinal evaluation, HCPAF should be used as a governance heuristic for interrogating AI-enabled admissions, not as proof that such systems will improve equity or trust. That honesty may prove to be the framework’s most important feature: at a moment when institutions face intense pressure to adopt AI, it offers a roadmap for doing so on educational rather than purely institutional terms.

Subject of Research: A human-centred AI and learning analytics governance framework for equitable personalised higher education admissions

Article Title: A human-centred artificial intelligence and learning analytics framework for equitable personalised higher education enrolment

Article References: Alenezi, A. F. (2026). A human-centred artificial intelligence and learning analytics framework for equitable personalised higher education enrolment. Discover Education, 5(1), Article 1084. https://doi.org/10.1007/s44217-026-02221-2

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02221-2

Keywords: human-centred AI, learning analytics, higher education admissions, educational equity, explainable AI, generative AI, learner agency, algorithmic bias, responsible AI governance, personalised learning, widening participation, student success