AI can write economics papers, but barely — it's the research ideas that fail, not the writing

Researchers compared 912 AI-generated economics papers to 41 published human papers and found the quality gap comes from two separate problems: AI generates weak research ideas (the bigger problem, 71% of the gap) and mediocre execution (29% of the gap). This means AI can handle the technical parts of research — statistical methods, robustness checks, writing — but struggles with the core task of academic work: asking interesting questions.

For three years, the economics field has been waiting to see whether AI systems could do actual research or just simulate it. This paper shows the answer is: AI can execute research competently but can't reliably generate novel ideas worth executing. The practical implication is blunt. If you're running a journal or a research institution, you cannot use AI as a replacement for the early stages of research — the part where a human decides what question to ask. You can use it to run analysis, check robustness, or write up findings. But the bottleneck is human creativity, not AI capability.