Devin Barbour

Research Fellow @Spar Research

London, GB

Standards for Uplift Studies\" in AI Safety EvaluationClarifying what qualifies as an uplift study in AI safety to build shared standards and comparability.Surveying existing methods from economics, psychology, machine learning, and evaluation organizations to map practices and gaps.Identifying recurring methodological pitfalls such as selection bias, regression to the mean, and poor generalization.Developing draft guidance on experimental design, data handling, validity checks, and transparent reporting.Proposing a reporting schema that includes uplift estimates, confidence intervals, assumptions, and generalization domains.Grounding methodological proposals in real-world practice through interviews and case studies with practitioners.