Measuring GenAI Code Usage and Standards

Measuring Code Quality in the Age of GenAI - Matt Van Itallie

Topic 6 of 11

Published Feb 10, 2025 · Clip 29:39–36:54 (7:15)

Matt explains how Sema detects GenAI code using a deep learning model trained on synthesized LLM code versus pre-GenAI open source code, refined with rules-based overrides. He lays out Sema's GenAI standards: the safe zone is 15-30% GenAI-originated code, and beyond volume, they distinguish pure versus blended GenAI code — more than 10% pure GenAI code triggers concern because GenAI hallucinates, lacks context, and is insecure. Proper protocols like non-training licenses can turn high usage into a strength.