Rod Pardo

Google automated the research paper. Google built a team of AI agents that writes a full research paper from a scientist's rough notes. The system is called PaperOrchestra. You give it a short summary of the idea and the raw experiment results, and it returns a finished paper formatted for a conference. Five AI agents split the work. One plans the outline, one draws the charts and diagrams, and one finds related research and checks that every cited paper exists. A fourth writes the sections, and a fifth revises the draft using feedback from a simulated peer review. The team tested it on 200 papers accepted at two top AI conferences. They reduced each paper to its idea and its results, then had the system write it again. ✅ It took about 40 minutes per paper on average. ✅ An AI reviewer gave its papers simulated acceptance rates of 84% and 81%. The real human-written papers got 86% and 94%. ✅ It cited about 46 to 48 papers per manuscript. Humans cited about 59, and other AI writers cited 9 to 14. ✅ Human judges preferred its literature reviews over other AI systems' by margins of 50% to 68%. ✅ Against the human-written literature reviews, it tied or won 43% of the time. The limits are in the paper too. A second AI reviewer gave it 48% and 22%, against 71% and 63% for the human papers. When 11 AI researchers compared full papers side by side, they preferred the human-written one 81% of the time. The acceptance rates come from AI reviewers, so I'd treat them as estimates. The system still needs a human to supply the idea and run the experiments, and the authors say humans stay accountable for what the paper claims.