Imperial Research Proposal Example: Applicant deciding MSc AI or MSc robotics (Score 93)

The governance literature on artificial intelligence has expanded rapidly since 2021, yet it largely treats 'AI systems' as a homogeneous category. Regulatory instruments — the EU AI Act, the UK AI Safety Institute's evaluation frameworks, ISO/IEC 42001 — apply risk-tiering logic designed primarily for software-based decision systems. Embodied robotic systems, which combine learned perception, physical actuation, and real-time autonomy, introduce distinct failure modes: a misclassified object in a warehouse sorting robot does not merely produce a wrong output but a physical consequence with liability, safety, and insurance implications that differ structurally from those of a language model generating a misleading sentence. This proposal asks: To what extent do existing AI governance frameworks adequately account for the technical and risk distinctions between software AI agents and embodied autonomous robotic systems, and what modifications to regulatory classification logic would produce more coherent governance outcomes? Two subsidiary questions follow. First, which technical properties of robotic systems — actuation, sensor-fusion latency, physical environment coupling — are absent from or misrepresented in regulatory risk taxonomies? Second, where classification ambiguity exists, what decision criteria do governance frameworks use, and are those criteria consistent across jurisdictions? The practical stakes are immediate. As autonomous systems enter logistics, healthcare, and public infrastructure, whether a system is governed as 'AI software' or 'machinery with embedded AI' determines which conformity assessment route applies, which liability regime attaches, and which regulator holds jurisdiction. Misclassifying the boundary has compounding downstream effects. Two bodies of scholarship bear on this question but have not been brought into productive dialogue. The AI governance literature — spanning algorithmic accountability, risk-based regulation, and the political economy of AI standards — has analysed how software-based AI systems should be classified, audited, and sanctioned. Scholars have examined the EU AI Act's prohibited-use categories, the opacity of high-risk system definitions, and conformity assessment for general-purpose AI models. This literature is technically informed about machine learning but treats physical embodiment as peripheral. The robotics regulation literature approaches governance from a product-safety and machinery-directive tradition, examining the revised EU Machinery Regulation (2023/1230), ISO 10218 standards for industrial robots, and whether autonomous mobile robots require new liability categories. This literature is precise about physical risk but has engaged only selectively with the algorithmic governance frameworks now applying to AI components embedded within robotic systems. The gap is structural: no systematic comparative analysis has examined how current AI governance instruments handle the boundary between software agents and embodied systems, nor has any study mapped classification criteria across the EU AI Act, the UK's pro-innovation framework, and the US Executive Order on AI (October 2023) against a consistent set of technical properties drawn from robotics engineering. This proposal addresses that gap. The study proceeds in three phases. Phase one: structured regulatory document analysis. The primary corpus comprises the EU AI Act (consolidated text, 2024), the UK AI Safety Institute's published evaluation methodology, ISO/IEC 42001:2023, the revised EU Machinery Regulation, and ISO 10218-1/-2. Each document will be coded against a technical property schema derived from robotics and AI engineering literature — covering actuation, sensor modality, real-time decision latency, physical environment coupling, and learning-during-deployment capability. The coding scheme will be developed iteratively using a pilot set of ten regulatory provisions, with inter-coder reliability checked via Cohen's kappa before full application. Document analysis is appropriate because the research question concerns the internal logic of classification criteria, not stakeholder perceptions. Phase two: comparative case analysis of three system archetypes — a large language model deployed as a customer-service agent (software-only), an autonomous mobile robot in a logistics warehouse (embodied, structured environment), and a surgical assistance robot under human supervision (embodied, safety-critical, human-in-the-loop). For each archetype, the analysis traces the regulatory classification pathway through each framework, identifying where criteria produce consistent outcomes, ambiguity, or contradictory assignments across jurisdictions. Case selection is purposive; the goal is to stress-test classification logic at boundaries where frameworks are most likely to diverge. Phase three synthesises findings into proposed classification criteria modifications, evaluated against three normative benchmarks from regulatory theory: coherence (internal consistency), proportionality (alignment between risk level and regulatory burden), and jurisdictional interoperability (compatibility with parallel frameworks). The output is a structured set of recommendations, not a draft regulation, keeping the contribution within master's dissertation scope. The choice to combine document analysis with structured case comparison is deliberate. Quantitative approaches suit neither the small population of regulatory instruments nor the qualitative variation of interest. Interview methods could supplement the analysis but are unnecessary for the primary question and would introduce access and ethics complexity disproportionate to the dissertation scope. All primary sources are publicly available or institutionally licensed. The EU AI Act, UK AISI documentation, ISO standards (accessible through library subscription), and the US Executive Order are open-access or covered by institutional agreements. No fieldwork, human participants, or proprietary datasets are required; the study falls outside human-subjects ethics review, to be confirmed with the departmental ethics officer at the outset. The principal feasibility risk is scope creep in regulatory document analysis. Mitigation: the technical property schema will be finalised and piloted in weeks one to four, and case archetypes are defined in advance rather than selected inductively. Provisional twelve-month timeline: months one to two — literature consolidation and coding scheme development; months three to four — pilot coding and inter-coder reliability check; months five to seven — full corpus coding and case pathway tracing; months eight to nine — synthesis and criteria modification drafting; months ten to eleven — writing and internal review; month twelve — submission and revision. If coding runs long, phase three can be scoped to two normative benchmarks without materially weakening the contribution. Imperial's Department of Computing and its affiliated policy and governance research activity sit at the intersection of technical systems analysis and regulatory design — precisely the disciplinary boundary this project occupies. The research draws on technical AI and robotics knowledge (the ability to critically assess system architecture descriptions in regulatory annexes) and on regulatory theory and comparative policy analysis. A BSc Computer Science background provides the technical grounding; the MSc programme provides the governance frameworks needed to operationalise the normative benchmarks in phase three. The project requires institutional library access for ISO standards, standard computing resources for qualitative data management, and supervision spanning AI systems and technology regulation. No laboratory access, specialist equipment, or external data agreements are needed. The bounded, document-based design is well-matched to resources available within a one-year taught master's programme. The expected contribution is modest and specific: a systematic mapping of classification criteria across three major governance frameworks against a consistent technical property schema, and defensible recommendations for where those criteria should be revised. This does not resolve the broader question of governing autonomous systems, but it provides a more precise analytical foundation for that debate than currently exists in the published literature.