Evolution of Programmers' Trust in Generative AI Programming Assistants

2025-09-16

AI Summary

This study investigates the dynamic evolution of programmers’ trust in generative AI programming assistants—specifically GitHub Copilot—to mitigate the dual challenges of overtrust (posing security risks) and undertrust (impairing productivity and user experience). Through a 10-day mixed-methods empirical study, we tracked trust development among 71 senior computer science students engaged in authentic software engineering tasks, integrating longitudinal surveys, in-depth interviews, and code artifact analysis. Results reveal an overall upward trajectory in trust, primarily driven by perceived correctness, contextual understanding capability, and foundational natural language processing proficiency. A key contribution is the first identification of the synergistic moderating effect of developers’ programming competence and tool-specific mental models on trust calibration. Based on these findings, we propose four actionable pedagogical recommendations to support students in achieving appropriately calibrated trust in AI-assisted programming.

Technology Category

• Philosophy and Ethics of AI: Safety, Robustness & Trustworthiness
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Humans and AI: Human-Aware Planning and Behavior Prediction

Application Category

• Economics, Online Markets and Human Computation: Trust and reliance of crowd workers and data experts on GenAI
• User Modeling, Personalization and Recommendation: Practical large-scale studies of user experience
• Social Networks and Social Media: Generative AI / large language models and their impact on social systems

Abstract

Motivation. Trust in generative AI programming assistants is a vital attitude that impacts how programmers use those programming assistants. Programmers that are over-trusting may be too reliant on their tools, leading to incorrect or vulnerable code; programmers that are under-trusting may avoid using tools that can improve their productivity and well-being. Methods. Since trust is a dynamic attitude that may change over time, this study aims to understand programmers' evolution of trust after immediate (one hour) and extended (10 days) use of GitHub Copilot. We collected survey data from 71 upper-division computer science students working on a legacy code base, representing a population that is about to enter the workforce. In this study, we quantitatively measure student trust levels and qualitatively uncover why student trust changes. Findings. Student trust, on average, increased over time. After completing a project with Copilot, however, students felt that Copilot requires a competent programmer to complete some tasks manually. Students mentioned that seeing Copilot's correctness, understanding how Copilot uses context from the code base, and learning some basics of natural language processing contributed to their elevated trust. Implications. Our study helps instructors and industry managers understand the factors that influence how students calibrate their trust with AI assistants. We make four pedagogical recommendations, which are that CS educators should 1) provide opportunities for students to work with Copilot on challenging software engineering tasks to calibrate their trust, 2) teach traditional skills of comprehending, debugging, and testing so students can verify output, 3) teach students about the basics of natural language processing, and 4) explicitly introduce and demonstrate the range of features available in Copilot.

Problem

• Studying evolution of programmers' trust in AI assistants
• Measuring how trust changes with extended Copilot usage
• Understanding factors influencing trust calibration in programmers

Innovation

• Measured trust evolution through longitudinal surveys
• Combined quantitative and qualitative trust analysis methods
• Studied GitHub Copilot usage on legacy code projects