AI-generated code quality and liability

Code security for software engineers

Topic 10 of 14

Published Nov 26, 2025 · Clip 47:43–51:36 (3:53)

Johannes describes Sonar's studies of the most popular LLMs (Claude, GPT-5, Llama, OpenCoder) and their 'personalities'—the kinds of issues and quality they produce. An interesting finding: GPT-5's reasoning mode reduces but doesn't eliminate security issues while producing more verbose, low-quality code that's harder to review and maintain, creating downstream security problems. The host connects this to the old adage that code is a liability and shares his practice of pausing vibe coding to refactor bloated output. Johannes notes a Stack Overflow survey where only 3% of developers trust AI-generated code, and agrees developers must keep code well structured and maintainable.