How well can AI actually write quantum computing code?

Translating Microsoft's quantum computing curriculum into Qiskit, researchers ran 16 AI models through 39,200 coding tasks to rigorously test quantum programming ability. The best models handle textbook algorithms like Grover's and Simon's with ease but fall apart when asked to encode a real problem into a quantum circuit — the genuinely hard part. Chain-of-thought prompting, often a reliable boost, actually hurt performance for most models except those explicitly trained to reason step-by-step.