Relationships Between Trust, Compliance, and Performance for Novice Programmers Using AI Code Generation
Research with 27 novice programmers finds no link between trust in AI assistants and actual compliance with their suggestions.
A new study from researchers at the University at Buffalo and the University of Illinois Urbana-Champaign challenges conventional wisdom about how novice programmers interact with AI coding assistants like GitHub Copilot. Published on arXiv, the research examined 27 novice programmers working under time pressure, measuring their subjective trust in the AI alongside objective metrics of coding performance and compliance with AI suggestions. The key, counterintuitive finding was that a programmer's level of trust did not determine how often they actually followed the AI's code recommendations.
Instead, the data revealed a different dynamic: programmers who performed well on tasks tended to comply more with the AI's suggestions, and this successful experience then boosted their trust in the tool afterward. This suggests that for novices, trust is an outcome of effective use, not a driver of it. The researchers conclude that focusing solely on building user trust in AI assistants may be less effective than designing training that directly promotes successful interaction outcomes, regardless of a user's initial trust level.
- Study of 27 novice programmers found no correlation between trust in AI coding assistants and compliance with their suggestions.
- Strong task performance was associated with higher AI compliance, which then led to increased trust in the tool.
- The findings suggest instructional design for AI tools should focus on promoting effective use rather than just building trust.
Why It Matters
This research could reshape how companies train developers on AI tools, shifting focus from building blind trust to fostering critical, effective collaboration.