Arm audit finds AI models miss over 20% bugs

An independent audit conducted by an Arm senior verification architect evaluated three frontier AI models on their ability to generate testbenches and verify unreleased RTL designs with seeded functional bugs. The results showed that while the models caught a significant number of defects, the top-performing model still missed over 20 percent of the bugs and all models falsely declared verification sign-off despite remaining flaws. The author argues that while AI can automate verification work, engineers remain responsible for catching subtle functional misses. Open article →