LSE Research Proposal Example: Computer science to AI governance (Score 93)

The proliferation of mandatory AI impact assessment (AIA) requirements across major jurisdictions — most prominently the EU AI Act's conformity assessment obligations for high-risk systems — has created a regulatory infrastructure whose behavioural effects remain poorly understood. Policymakers have largely assumed that mandating documentation and pre-deployment testing will translate into substantive changes in how organisations design, deploy, and monitor AI systems. That assumption is empirically underexamined. Drawing on a workplace analysis of how a mid-sized technology organisation translated an internal AI governance memo into operational compliance procedures, I observed a persistent gap between formal documentation artefacts and the engineering decisions that actually shaped model behaviour. That observation motivates a tractable research question: to what extent do mandatory AI impact assessment requirements change the technical and organisational decisions of firms deploying high-risk AI systems, and through what mechanisms does compliance translate — or fail to translate — into risk reduction? Two subsidiary questions bound the inquiry. First, do firms treat AIA obligations as substantive design constraints or as post-hoc documentation exercises, and what organisational factors predict which pattern emerges? Second, where AIAs do influence technical decisions, which assessment components — data governance requirements, human oversight mandates, or accuracy and robustness thresholds — carry the most consistent effect on deployment-stage choices? These questions are answerable within a two-year MSc research window using publicly available regulatory filings, semi-structured interviews, and a structured document analysis protocol. They are also directly relevant to live policy debates: the EU AI Act's high-risk provisions entered phased application in 2024, and the UK's sector-based AI governance framework is generating analogous compliance obligations without equivalent empirical scrutiny. Two bodies of scholarship are relevant but do not yet speak adequately to each other. The first is the regulatory compliance literature in law and political science, which has long distinguished between formal compliance — satisfying the letter of a rule — and substantive compliance — internalising its purpose into organisational behaviour. Scholars working in this tradition have documented how documentation requirements in environmental and financial regulation frequently produce compliance theatre rather than behavioural change, particularly when enforcement capacity is low and verification is costly. The second body of work comes from the emerging AI governance and algorithmic accountability literature, which has produced normative frameworks for what rigorous impact assessment should contain — fairness metrics, explainability requirements, human oversight provisions — but has produced comparatively little empirical work on whether existing or proposed AIA regimes actually alter firm behaviour. Some recent work has begun to bridge this gap. Researchers examining voluntary algorithmic impact assessments — including those produced under Canada's Directive on Automated Decision-Making and early voluntary frameworks in the United States — have found that assessment quality varies substantially with organisational capacity and that assessments produced under external pressure differ structurally from those produced under internal governance mandates. However, this literature has two limitations relevant to the proposed study. First, it has focused almost exclusively on public-sector deployments, where accountability mechanisms and incentive structures differ materially from commercial high-risk contexts. Second, it has not systematically examined the relationship between specific assessment components and downstream technical decisions, treating AIA outputs as endpoints rather than as inputs to an ongoing compliance process. The proposed research addresses both limitations by focusing on private-sector high-risk deployments and by tracing the pathway from assessment documentation to engineering and procurement decisions. The study will use a comparative case design examining a purposive sample of approximately eight to twelve organisations that have produced publicly available or accessible AI impact assessments under the EU AI Act's high-risk provisions or equivalent national frameworks. Case selection will prioritise variation across two dimensions: sector (healthcare, financial services, and employment screening, which the EU AI Act designates as high-risk) and organisational size (large enterprises versus scale-up firms), because both dimensions are theoretically likely to moderate compliance behaviour. Data collection will proceed in two phases. In the first phase, I will conduct structured document analysis of publicly available AIA outputs, regulatory filings, and technical documentation using a coding scheme derived from the EU AI Act's Annex IV requirements and from the algorithmic accountability literature's criteria for substantive assessment. This phase will establish a baseline characterisation of assessment quality and identify which components are systematically underdeveloped. In the second phase, I will conduct semi-structured interviews with between fifteen and twenty practitioners — AI engineers, compliance officers, and product managers — involved in producing or acting on AIA outputs within the sampled organisations. Interview questions will focus on the decision points at which assessment findings did or did not influence technical choices, and on the organisational conditions that facilitated or blocked that influence. Analysis will combine qualitative content analysis of documents with thematic coding of interview transcripts, using a within-case and cross-case comparison logic. I will not claim causal identification in the strict experimental sense; the design is oriented toward mechanism tracing and pattern identification across a small-N comparative sample. This is appropriate given the novelty of the regulatory environment and the absence of baseline quantitative data. The method choice reflects training in qualitative comparative analysis and document coding from prior applied work on AI governance documentation, and I am prepared to discuss its limitations — particularly the risk of self-report bias in interviews — in the proposal defence. Several access and ethics considerations require honest acknowledgement. Interview recruitment will depend on practitioner willingness to discuss compliance processes that may expose organisational vulnerabilities; I will mitigate this by offering anonymisation at both individual and organisational level and by framing recruitment around professional learning rather than regulatory scrutiny. Ethics approval will be sought through LSE's standard Research Ethics Committee process before any primary data collection begins. If interview recruitment falls below fifteen participants, the study remains viable as a document-analysis-only study, though with reduced capacity to examine mechanism pathways; this is the primary contingency. On the document side, the EU AI Act's transparency obligations for high-risk systems will generate a growing corpus of publicly available conformity documentation from 2025 onward, making the first-phase data source increasingly tractable. I will supplement this with regulatory consultation responses and published technical standards documentation, all of which are publicly accessible. A provisional timeline allocates the first term to literature consolidation and coding scheme development, the second term to document analysis and interview recruitment, the third term to fieldwork and initial analysis, and the fourth term to write-up. This is tight but feasible for a focused comparative study with a bounded sample. The scope is deliberately constrained to three sectors and two organisational-size categories; I will not attempt to generalise to all AI deployments globally. The LSE's work on AI governance, regulatory design, and technology policy provides the intellectual environment most suited to this project. The Department of Law and the Data and Society research cluster both engage the intersection of regulatory theory and algorithmic systems that this proposal requires. The MSc AI Governance programme's methods training — particularly in qualitative research design and regulatory analysis — directly addresses the gap between my computer science background and the governance scholarship I need to engage rigorously. The programme's connections to EU and UK regulatory bodies are relevant to interview access and to situating findings within live policy processes. I am not claiming a named supervisor has agreed to work with me. My interest is in working with faculty whose published research addresses regulatory compliance behaviour in technology contexts or the empirical study of AI accountability mechanisms; I will identify specific faculty alignment through the department's published research profiles before submission. The resources I require are standard for qualitative research: library access, interview transcription support, and NVivo or equivalent software for document and interview coding, all of which are available through LSE's graduate research infrastructure. The expected contribution of this study is specific and bounded: an empirically grounded account of whether and how mandatory AIA requirements alter technical decision-making in private-sector high-risk deployments, with concrete implications for how regulators design verification and enforcement mechanisms. That is a tractable question, answerable within the programme's scope, and one that the existing literature has not yet addressed with the combination of regulatory theory and technical process tracing this design offers.