GPT-5.6 Sol disproves 20-year-old statistics conjecture in 90 minutes: the BH procedure is not foolproof

## GPT-5.6 Sol disproves a 20-year-old statistics conjecture in 90 minutes Edgar Dobriban, an associate professor at the Wharton School of the University of Pennsylvania, used GPT-5.6 Sol Pro to disprove a conjecture that had stood for about 20 years in the field of multiple hypothesis testing. He published the result as an arXiv preprint and listed himself as the sole author. **Key points** - **What was disproved**: The Benjamini-Hochberg (BH) procedure is a foundational method for controlling the false discovery rate (FDR), and its original 1995 paper has been cited more than 130,000 times. The conjecture held that BH controls the FDR for "correlated two-sided Gaussian tests" under any correlation structure. The AI showed that this is not the case. - **90 minutes vs. 20 hours**: In a single attempt of about 90 minutes, the model produced a complete counterexample, a proof and a machine-checkable numerical certificate. Dobriban revealed that GPT-5.5, even with multiple agents iterating in parallel for more than 20 hours, failed to solve it, a vivid illustration of the generational gap. - **Academic norms**: Dobriban stated explicitly in the preprint that the proof was obtained by GPT-5.6 Pro and listed himself as the sole author. A new paradigm of "AI produces the proof, humans take responsibility" is taking shape. > Source: [read the original](https://mlq.ai/news/openais-gpt-56-sol-disproves-20-year-old-statistics-conjecture-in-90-minutes/). This text is a summary for reference only.