Authors: Sadra Sabouri, Philipp Eibl, Xinyi Zhou, Morteza Ziyadi, Nenad Medvidovic, Lars Lindemann, Souti Chattopadhyay Venue: IEEE/ACM International Conference on Software Engineering (ICSE), 2025 — accepted January 2025 Method: survey of 29 developers, plus observation of 10 developers coding with AI
Summary
The paper starts from a question that sounds simple and isn’t: what do developers actually mean when they say they trust a code suggestion? Through surveys and through watching people code with AI, the study pins down how developers define trust in AI-generated code, what they weigh when they grant it, and what happens to it afterwards.
Two properties dominate the moment of decision. Developers judge a suggestion mostly on whether they can read it — comprehensibility — and on whether it looks correct. But that judgment is not where trust ends. Developers keep changing their minds after the fact, revising decisions they had already made, and in the end only 52% of AI suggestions survive to be accepted. Trust here is a moving quantity rather than a verdict issued once.
The study closes with four validated guidelines for the people building these tools, aimed at developer–AI collaboration rather than at the model itself.
Links
Related
Thread: Human-AI Interaction in Knowledge Work — first study in the thread; partially funded by the Amazon AGI center. Nenad Medvidovic — co-author.