How people who already have a job to do trust, oversee, and integrate AI inside the tools they work in. The thread began with software developers and now reaches beyond them: the same questions about trust, oversight, and where an agent should be allowed to act turn out to apply to a spreadsheet user as much as to a programmer. Knowledge work, rather than software engineering, is the level the findings generalise at.
- Trust — surveyed 29 and observed 10 developers on how they define, evaluate, and evolve trust in AI-generated code → Trust Dynamics in AI-Assisted Development: Definitions, Factors, and Implications
- Cognitive bias — observational study of 36 developers producing a taxonomy of 15 bias categories covering 90 biases → Cognitive Biases in LLM-Assisted Software Development
- Project-level agents — 16 developers using Cursor-like tools, identifying integration challenges and opportunities → Exploring the Challenges and Opportunities of AI-assisted Codebase Generation
- Interactive code querying — an educational prototype letting programmers query source code, evaluated with 20 software engineers → Synthesizing Program Analyzers to Help Programmers Answer Questions About Code
- Agents in spreadsheets — how users audit and retain control over AI agents acting inside a spreadsheet, taking the oversight question to a population that does not write code → Auditing and Controlling AI Agent Actions in Spreadsheets
Partially funded by the Amazon AGI Center.
Related
USC — Adaptive Computing Experience (ACE) Lab — the lab. Thread: AI Aid for Decision Making — the same oversight question where the output is advice rather than an artifact.