Coding assistants are now generating code that fails to perform as intended, but which on the surface seems to run successfully, avoiding syntax errors or obvious crashes. Notably, GPT 5 performed worse than GPT 4 in testing. spectrum.ieee.org/ai-coding-de...