Oxford Internet InstituteAI Evaluation
I’m Louis, an MSc candidate at the Oxford Internet Institute. My research asks whether the benchmark scores used to judge AI systems can be trusted, and follows those scores into the things they decide: prices, wages, and rules.
ice hockeyalpine ski racingpiano
Benchmarks are the exams AI models sit, and an exam is only useful if it is fair, consistent, and testing the right thing. Much of this research applies the statistics of human testing — psychometricsThe statistical field behind standardised testing: how to tell whether a test is reliable, comparable, and measuring what it claims. — to the exams we set for machines, and then follows the scores out into prices, wages and rules.
Twelve benchmarks that claim to measure different skills mostly track a single underlying factor.
421 modelsHash-pinned frontier model configurations in the frozen twelve-benchmark battery. 74.5% one factorShare of common variance on the first factor, exploratory factor analysis. As reported in the manuscript.
A price index for AI that holds capability constant — the way statistical agencies price computers.
782 modelsModels in the frozen 31-month price panel, each observation with source URL and archive timestamp. 87% hidden declineShare of the measured decline in the price of capability that standard matched-model methods cannot see, because it arrives through new models rather than price cuts on old ones.
Cross-lab standards for what a field must show before an evaluation counts as measurement. I write the validity and evidentiary-standards sections.
When an AI joins a group decision, its institutional role matters more than its accuracy.
1.1M simulationsRuns of the agent-based model across authority structures and information conditions.
The reliability statistics everyone borrows for AI judges rest on assumptions those judges break.
Six years of MMLU, the most-cited AI benchmark: the score depends on how the test is run, and the instability reaches the leaderboards built on it.
Does being measured as automatable move your wage? UK risk scores linked to pay records, occupation by year.
If firms must disclose AI risk, disclosure becomes a signal: audits, costs and penalties as a game, under the EU AI Act and the SEC’s proposed rules.
Also: adversarial CAPTCHAs that humans pass and AI agents fail, with UCL Computer Science and Holistic AI · a framework for when neural data should inform welfare policy arXiv
Departmental Associate · UCL Science & Technology Studies
Designing a new module on AI, digital labour and the future of work, funded by UCL MAPS. Supervised by Dr Joanna Octavia
Summer Researcher · LSE Department of Statistics
AI measurement and the factor structure of frontier evaluation, with Prof Marcos Barreto
Researcher · EvalEval Coalition
Cross-lab evaluation standards with Hugging Face, Edinburgh and EleutherAI; writing the validity sections of the Science of Evaluations paper
Teaching Assistant & Course Developer · UCL STS
Responsible Innovation in Practice and Governance of Emerging Technologies
Student AI Researcher · Holistic AI
Adversarial robustness of LLM and vision–language agents
Student Researcher · UCL Computer Science
Agent-based modelling of AI authority in group decisions, with Prof Maarten Speekenbrink — published at CHI 2026
Research Contributor · Institute of Economic Affairs
UK graduate premium, contributing to Julian Jessop’s analysis of administrative earnings data
Research Contributor · UCL Institute for Global Prosperity
AI & Youth, with Honorary Professor Noreena Hertz
For research, collaboration, or data and code requests: