Towards a Science of Causal Interpretability in Deep Learning for Software Engineering

Neural code models lack transparency about cause-and-effect relationships: While AI systems for software engineering perform well, they don't clearly explain why they make predictions—only what correlations they find. This distinction matters because understanding causation is essential for tasks requiring intervention and change analysis.

A new method called DoCode uses causal inference to explain code predictions: Rather than just identifying correlations, DoCode traces through programming language properties to provide explanations grounded in how code actually works, helping reduce misleading correlations that don't reflect real causal relationships.

AI code models are sensitive to syntax changes in ways that may not reflect true understanding: The research found that neural code models' predictions shift when code syntax changes, suggesting they may rely on surface-level patterns rather than deeper programming concepts—an important consideration for trusting these systems in real software engineering tasks.

Building trust in AI for software engineering requires moving beyond correlation to causation: The dissertation emphasizes that for practitioners to confidently use neural code models, they need explanations that show causal relationships, not just statistical associations, particularly when these systems are used to make decisions or automate changes.