Five ChatGPT-based AIs form council to debate USMLE questions. Council scores 97%, 93%, 94% on three USMLE sections. Team outperforms single AI models. No special medical training or data used. Tested on 325 basic and clinical questions. AIs compare answers, justify, reach consensus. Mediator resolves disagreements by summarizing arguments. Second discussion round fixes 53% of errors. Consensus beats individual AI potential. Provides first proof of AI self-correction via structured dialog. Collective AI intelligence tops solo performance. Method sets new benchmark for AI in medicine. Future potential to aid doctor decisions.