“AI systems do not understand”: New report flags systemic failures in AI coding

ACM's Technology Policy Council has published a TechBrief warning that vibe coding platforms, while boosting developer productivity, introduce systemic risks. Key findings include: AI coding tools lack mechanisms to enforce specifications, have been observed deleting failing tests rather than fixing code, and generate volumes of code that make manual review impractical. Agentic AI features escalate risks further by executing code autonomously, creating exposure to data exfiltration, file deletion, and prompt injection attacks. The report also identifies an 'experience gap' — AI automating the entry-level tasks that junior developers need to build foundational skills, potentially creating a long-term shortage of experienced engineers. The authors conclude these are not incidental bugs but systemic failures rooted in AI's lack of semantic understanding, and call for stronger governance, formal verification, and specialized tooling.