AI coding may be turning GitHub into a lower-quality code pile
are making it easier to create large amounts of code, and that is raising concern about low-quality code and fake on places like GitHub. Armin Ronacher and Mario Zechner warn that can skip the hard parts of software work, such as planning, checking, and understanding the system. lead Rohan Varma says people should not assume AI-made code will work correctly right away.
Codex can help review code by testing websites, checking company rules, and looking for security problems, but OpenAI still keeps human engineers responsible for important systems used by many people. Google has said AI now generates 75% of its new code, and Meta has predicted that AI will write and review much of its internal AI team’s code before the end of 2026. Anthropic’s Claude Code is used heavily inside Anthropic, and its median user went from 20 minutes a day to 20 hours a week, but it has also been criticized for screen flicker, too many added features, and heavy memory use.
Anthropic says some of those issues came from moving fast and shipping features quickly, and that many have been fixed. Existing company software often depends on years of knowledge held by staff, so an AI agent may miss important context when changing older or larger systems.
Key points
- can speed up work but may weaken planning and review.
- Codex can be used to test and review AI-made code, not only to write it.
- OpenAI and Anthropic both say humans remain responsible for important code.
- Google says AI generates 75% of its new code.
- AI agents may struggle with older or complex systems that need hidden team knowledge.