UC Santa Barbara

AI systems increasingly provide advice and perform work autonomously. People must judge the evidence behind that advice and define the objectives and evaluation procedures for delegated work. This dissertation studies how people can understand AI behavior and design environments that guide autonomous work and make progress on important problems independently verifiable.We first develop methods that connect model explanations to inspectable evidence and study how decision-support interfaces affect human judgment. In the textual-explanation condition of our decision study, perceived reliability increases without a detectable improvement in accuracy. These findings distinguish access to evidence from its effective use in decisions.We next examine how task conditions and system configurations shape capability. Controlled programming evaluations show that relative model performance varies across task structures. Our visualization workflow organizes tools and feedback to support revision, while trajectory audits reveal inconsistencies between observations and subsequent actions. These studies connect capability evaluation to the task, the configured system, and the execution history. We study improvement capability, the ability to produce independently verified gains under stated task conditions and resource limits. AutoLab turns research and engineering problems contributed by experts into executable environments with working baselines, explicit constraints, local feedback, and independent evaluation. Agents achieve verified implementation gains while their model weights remain fixed, but differ in how they use experiments and resources. Some runs with local progress receive no credit under independent evaluation, and changes in execution configuration alter performance and model rankings.Together, these studies provide methods for evaluating AI behavior and making the requirements for meaningful progress explicit in environments for autonomous work. The findings also motivate a research agenda for human–AI co-evolution: how AI systems reshape human knowledge and practices, and how human experience informs subsequent system design.