Oxford Research Proposal Example: Law graduate deciding LLM or MSc computer science (Score 93)

The EU AI Act (2024) imposes transparency and explainability obligations on providers of high-risk AI systems, yet delegates the technical content of those obligations almost entirely to harmonised standards that do not yet exist. The legal duty to explain an automated decision is formally in force, but the technical threshold at which an explanation is deemed legally adequate remains undefined. Courts, regulators, and affected individuals therefore operate in a normative vacuum that neither the legal nor the computer science literature has adequately mapped. This proposal asks: Under what conditions do current machine-learning explainability methods — specifically post-hoc attribution techniques such as SHAP and LIME — satisfy the transparency requirements imposed on high-risk AI systems by the EU AI Act, and where do those methods fall structurally short of legal adequacy? Two subsidiary questions follow. First, how do the Act's requirements compare with the interpretability standards implicitly assumed by the GDPR's Article 22 right not to be subject to solely automated decisions? Second, what minimum technical specification would a harmonised standard need to adopt to close the gap between current explainability practice and legal compliance? The question is bounded: it concerns a defined statutory instrument, a defined class of AI systems (high-risk, Annex III), and a defined set of explainability methods. It is not a general inquiry into AI ethics or a survey of global regulatory approaches. Two bodies of scholarship bear on this question but have not been brought into productive contact. The legal literature on algorithmic accountability — represented by Wachter, Mittelstadt, and Russell on counterfactual explanations, and Goodman and Flaxman on GDPR's right to explanation — has focused on whether a legal right to explanation exists and what its normative justification might be. This literature is largely pre-Act, written in anticipation of regulation rather than in response to enacted text, and tends to treat explainability as a binary entitlement rather than a technical specification problem. The computer science literature on interpretable machine learning — surveyed in Molnar's open textbook and in empirical work comparing SHAP, LIME, and attention-based methods — addresses fidelity, stability, and computational cost but rarely asks whether these properties map onto legal standards. The few bridging papers (for instance, Ustun, Spangher, and Liu on actionable recourse) operate in the context of individual fairness rather than regulatory compliance. The gap is therefore methodological and doctrinal simultaneously: no systematic framework translates the Act's legal language — "appropriate level of transparency," "human oversight," "meaningful information" — into verifiable technical criteria. Without such a framework, compliance assessments by national market surveillance authorities will be inconsistent and the Act's enforcement mechanism weakened from the outset. This proposal does not claim to resolve the gap definitively within a one-year MSc. It aims to produce a structured analytical framework mapping the mismatch and identifying the minimum specification conditions a harmonised standard would need to satisfy. The research proceeds in three phases. Phase one (months one to three): doctrinal analysis. I will conduct a close reading of the EU AI Act's transparency provisions — Articles 13, 14, and 50 — alongside contextualising recitals, the Commission's explanatory memorandum, and parliamentary amendments. I will compare these with the GDPR Article 22 framework and relevant CJEU and national supervisory authority decisions. The output is a structured legal specification: conditions an explanation must satisfy to be legally adequate under the Act. Phase two (months three to six): technical analysis. I will apply SHAP and LIME to two publicly available, high-risk-adjacent datasets — the COMPAS recidivism dataset and the UCI Adult Income dataset — used extensively in the fairness and interpretability literature and raising no novel access or consent issues. I will evaluate outputs against the phase-one legal specification using four criteria derived from the Act's text: fidelity to the model's actual decision process, stability across similar inputs, actionability for the affected individual, and auditability for a third-party regulator. These datasets are not claimed to represent real deployment contexts; they serve as analytical test cases to stress-test the legal specification against known technical behaviour. Phase three (months six to nine): synthesis. I will draft minimum technical specification conditions and assess whether any existing harmonised standard proposal — including CEN-CENELEC JTC 21 working documents — approaches adequacy. The dissertation concludes with a gap analysis and concrete drafting recommendations addressed to the standardisation process. The choice to combine doctrinal legal analysis with applied technical evaluation is deliberate. I am not conducting empirical social research or interviewing regulators. The scope is analytical: legal text against technical output. This keeps the project feasible within twelve months while producing a contribution that neither a pure lawyer nor a pure computer scientist would be positioned to make alone. All primary legal materials are publicly available: the Act, the GDPR, CJEU judgments, and the Commission's standardisation mandate. The COMPAS and UCI Adult Income datasets are freely available under standard academic licences and contain no personally identifiable information in the versions used for interpretability research. No new data collection is required, and no ethics approval beyond standard institutional review for desk-based research is anticipated — though I will confirm this with the relevant Oxford committee at the outset. The principal feasibility risk is technical depth. Implementing SHAP and LIME requires competence in Python and familiarity with the scikit-learn and SHAP libraries. I have developed this competence through independent study and a structured self-directed project analysing LLM output consistency, which provided working experience with Python-based NLP pipelines. The technical phase requires evaluation rather than model development; the libraries are well-documented and the datasets pre-processed in standard formats. A secondary risk is that the harmonised standards process may advance faster than anticipated, rendering parts of the gap analysis obsolete. This is manageable: the doctrinal and methodological framework would retain value even if specific standards are published during the research period, since the framework itself is the contribution. Timeline: doctrinal analysis and legal specification (months one to three); technical implementation and evaluation (months three to six); synthesis, gap analysis, and drafting recommendations (months six to nine); dissertation writing and revision (months nine to twelve). Oxford's Internet Institute and the Faculty of Law both maintain active research programmes on platform regulation, algorithmic accountability, and AI governance. The OII's published work on automated decision systems and the Law Faculty's engagement with EU digital regulation provide the scholarly environment in which this project sits. I am not claiming any supervisor commitment; I identify publicly documented research clusters — including the OII's data and society programme and the Centre for Technology and Global Affairs — as the appropriate intellectual context. The project requires access to legal databases (Westlaw, EUR-Lex), standard computing resources for Python-based analysis, and library access to the interpretable machine learning literature — all standard MSc-level resources. No fieldwork, travel, or specialised laboratory access is required. The expected contribution is modest and specific: a structured analytical framework for assessing the legal adequacy of explainability methods under the EU AI Act, tested against two benchmark datasets, with drafting recommendations for the harmonised standards process. This is a tractable twelve-month research problem at the intersection of legal doctrine and technical evaluation — a boundary the MSc in Technology Governance is positioned to support.