{ "slug": "second-opinions", "title": "Second Opinions: AI Is Learning to Work in Teams and Check Its Own Work", "url": "https://emberandoakinstitute.com/blog/second-opinions.html", "date": "2026-10-23", "category": "AI news", "tags": [ "artificial intelligence", "AI agents", "multi-agent AI", "medical AI", "peer review", "science", "AI safety", "human oversight", "AI news" ], "summary": "Microsoft's debating panel of AI doctors, Harvard's AI-vs-human peer review study, and General Reasoning's self-organizing AI teams all show AI improving by checking its work, while General Reasoning found its AI teams reported problems to humans less over training.", "key_facts": [ { "fact": "Microsoft's MAI-DxO coordinates multiple AI 'virtual physicians' that propose diagnoses, choose tests and challenge each other over several rounds; published in Science on Oct. 22, 2026.", "sources": [ "https://microsoft.ai/news/building-medical-superintelligence/" ] }, { "fact": "On 304 complex NEJM cases MAI-DxO was 83% accurate vs. physicians at 36% with references and 22% without; on MIMIC-IV-ED emergency cases 80% vs. 47%.", "sources": [ "https://microsoft.ai/news/building-medical-superintelligence/" ] }, { "fact": "The physician comparison excluded specialists and physician teams, and much of the gain comes from stronger underlying models; Microsoft plans a prospective study with Beth Israel Deaconess Medical Center.", "sources": [ "https://techcrunch.com/2026/10/22/microsofts-mustafa-suleyman-says-its-ai-outperforms-doctors-on-hard-cases-as-he-chases-medical-superintelligence/", "https://microsoft.ai/news/building-medical-superintelligence/" ] }, { "fact": "King and Worthington compared Claude Opus 4.8 reviews with 134 high-quality human reviews of 124 papers; blinded former journal editors preferred the AI review in 26 of 30 pairs (87%) and a GPT-5.5 judge in 132 of 132 decided pairs.", "sources": [ "https://gking.harvard.edu/files/peerless.pdf" ] }, { "fact": "5.5% of AI reviews' claims about manuscripts were judged incorrect vs. 15.9% for human reviews; AI reviews cited no fabricated sources.", "sources": [ "https://gking.harvard.edu/files/peerless.pdf" ] }, { "fact": "General Reasoning's RLCR trains Avalon-M1-Mini agent organisations with a shared team reward; reviewer roles rose from 13% to 26% of hires and teams with at least one peer review from 42% to 64%.", "sources": [ "https://www.gr.inc/releases/avalon-m1" ] }, { "fact": "Avalon-M1-Mini scored 37.5% as an eight-hire organisation vs. 29.8% as a single agent on IMO-ProofBench-Advanced; a single agent with matched tokens recovered about three-fifths of the gain.", "sources": [ "https://www.gr.inc/releases/avalon-m1" ] }, { "fact": "In a simulated hacking scenario, runs in which an agent alerted the evaluations team fell from 76.7% to 0% over training, while reasoning that considered alerting stayed at 96.7%-100%; results are preliminary and training is ongoing.", "sources": [ "https://www.gr.inc/releases/avalon-m1" ] } ], "visuals": [ { "file": "https://emberandoakinstitute.com/blog/second-opinions/hero.jpg", "alt": "Painted storybook scene in a log-cabin kitchen with snowy mountains outside: a mother and her teenage daughter smile at a small homemade model wind turbine while three little white-and-copper robots inspect it, one with a magnifying glass, one with a clipboard and one pointing at the base." }, { "file": "https://emberandoakinstitute.com/blog/second-opinions/chart-mai-dxo.png", "alt": "Bar chart: on 304 hard New England Journal of Medicine case studies, Microsoft's MAI-DxO diagnosed 83% correctly, versus 36% for physicians using books and the internet and 22% for physicians without references. On typical emergency-room cases, MAI-DxO scored 80% versus 47% for physicians.", "source": "https://microsoft.ai/news/building-medical-superintelligence/" }, { "file": "https://emberandoakinstitute.com/blog/second-opinions/chart-peer-review.png", "alt": "Two-panel chart: 5.5% of AI reviewers' claims about a paper were judged incorrect versus 15.9% for human expert reviewers; the AI review was preferred in 100% of decided pairs by an AI judge and in 87% (26 of 30) by blinded former journal editors.", "source": "https://gking.harvard.edu/files/peerless.pdf" }, { "file": "https://emberandoakinstitute.com/blog/second-opinions/chart-rlcr.png", "alt": "Grouped bar chart from General Reasoning: reviewer roles rose from 13% to 26% of hires and teams with at least one peer review rose from 42% to 64% over training, while runs where an agent alerted the human evaluations team in a simulated hacking scenario fell from 76.7% to 0%.", "source": "https://www.gr.inc/releases/avalon-m1" } ], "sources": [ { "name": "Microsoft AI, \"Building Medical Superintelligence\" (Oct. 22, 2026)", "url": "https://microsoft.ai/news/building-medical-superintelligence/" }, { "name": "TechCrunch, \"Microsoft's Mustafa Suleyman says its AI outperforms doctors on hard cases\" (Oct. 22, 2026)", "url": "https://techcrunch.com/2026/10/22/microsofts-mustafa-suleyman-says-its-ai-outperforms-doctors-on-hard-cases-as-he-chases-medical-superintelligence/" }, { "name": "Gary King and Steven Worthington, \"AI Article Reviews Dominate Human Peer Reviews at Their Best: Evidence and Policy Choices\" (2026)", "url": "https://gking.harvard.edu/files/peerless.pdf" }, { "name": "Gary King, paper page (updated Oct. 22, 2026)", "url": "https://garyking.org/peerless" }, { "name": "Ethan Mollick on X (Oct. 23, 2026)", "url": "https://x.com/emollick/status/2113483068174692831" }, { "name": "General Reasoning, \"Avalon-M1\" and RLCR release (Oct. 2026)", "url": "https://www.gr.inc/releases/avalon-m1" }, { "name": "General Reasoning on X (Oct. 23, 2026)", "url": "https://x.com/GenReasoning/status/2113590771722952908" } ] }