RDEL #43: How do software engineers build trust in AI-powered tools?
The AI-powered tooling landscape is rapidly evolving. This week we look at research on how engineers build trust in these tools, specifically for code generation.
Welcome back to Research-Driven Engineering Leadership. Each week, we pose an interesting topic in engineering leadership, and apply the latest research in the field to drive to an answer.
The ecosystem of AI-powered code generation tools has exploded over the last year, which means nearly every software engineer has had to evaluate whether these tools are right for them. But where do they go to build trust in these tools? This week, we dive into the research on how developers build trust in AI-powered code generation tooling.
The context
In the last couple of years, the advancement of generative AI models has significantly impacted software development. AI-powered code generation tools such as GitHub Copilot, Tabnine, and others have emerged, improving developer productivity by automating numerous parts of the coding process.
The adoption and effective use of these tools depend heavily on the developers' trust in their capabilities and reliability. There are numerous ways to build this trust - for example, online communities play at least some role in shaping this trust by providing shared experiences and evaluation signals that help developers make informed decisions about utilizing these AI tools. This paper dives into the specific mechanisms by which software engineers build that trust when they evaluate new AI-powered tooling for code generation.
The research
Researchers conducted a two-part study. In the first part, the researchers interviewed 17 developers to understand how online communities help them build trust in AI code generation tools. In the second part, the researchers developed design probes based on the insights from the interviews and presented them to 11 developers to explore how community features could be integrated into AI tools to support trust-building.
Their research led researchers to discover that trust is built through two primary mechanisms: community-curated experiences and community-generated evaluation signals.
Community-curated experiences: Developers build trust through vivid descriptions of AI interactions, realistic programming tasks, diverse use cases, and detailed setup and dependency information. These community-shared experiences help developers form accurate mental models of AI capabilities.
Community-Generated Evaluation Signals: Signals like user votes, comments, and engagement statistics from online communities help developers assess the quality and reliability of AI-generated code suggestions. These signals support developers in making informed trust judgments.
The application
The findings from this study show a common pattern: software engineers build trust from recommendations in their community. When it comes to evaluating a new AI-powered tool, the experience of other engineers will weigh heavily on the evaluation to adopt a new tool. (Note: a previous edition evaluated how developers adopt new tools more broadly).
For managers looking to adopt new AI-powered tools, consider the community support behind a tool as a signal of whether it is widely used in the community. The tooling landscape moves quickly, but a trustworthy tool will likely have online discourse about its effectiveness, which can help your team decide whether it makes sense for their use cases. Additionally, foster a culture within the team that encourages sharing experiences and feedback about AI tools. This can further support collective sense-making, helping teammates build appropriate trust and make better use of AI-powered tools.
—
We hope this research helps make trust-building in AI-tooling a bit more methodical. Happy research Monday!
Lizzie