An Empirical Study of Knowledge Transfer in AI Pair Programming [pdf] (www.se.cs.uni-saarland.de)

Researchers at Saarland University and Siemens ran an empirical study comparing knowledge transfer in traditional human–human pair programming versus human–AI pair programming using GitHub Copilot. Nineteen developers either worked in pairs or individually with Copilot on a non-trivial programming task that required algorithm design and integration. The team unified and extended prior knowledge-transfer frameworks (Zieris & Prechelt; Kuttal et al.) and used a semi‑automated evaluation pipeline to identify and classify transfer episodes across both settings. They found that successful knowledge-transfer episodes occur at similar frequencies and cover overlapping topical categories in both human and AI-assisted settings. Key behavioral differences emerged: human pairs produced more frequent (and sometimes distracting) exchanges, while Copilot interactions were more focused—but developers accepted Copilot’s suggestions with less scrutiny. Copilot also provided subtle, useful reminders (e.g., missing DB commits) that humans might overlook. Technical caveats include that the study used Copilot backed by CODEX at the time. Implications: AI assistants can reproduce many knowledge-transfer benefits but risk reduced critical engagement; tool and workflow design should encourage explanations, challenge suggestions, and hybrid human–AI setups to preserve learning and code quality while boosting productivity.