I had OpenAI extract coefficients and standard errors from 3500 estimates across 651 “AI automated manuscripts” to see if they were p-hacking. And in scientific terms, they are p-hacking the hell out of these papers — even without ANY motivation to get published or instructions to produce anything in particular. Read the whole thing — real data, real policy variation, real estimators, real AI generated start to finish writing, real bunching at t statistics of 1.96, and particularly severe for IV — just like humans!