AI may outdiagnose doctors, but silent data omissions are the real danger
This commentary, translated from Unite.AI, argues that LLMs already beat an individual doctor on some cognitive tasks: it cites a 2024 randomized trial with 50 doctors (GPT-4 group 76 vs 74 for traditional resources, not significant, while GPT-4 alone scored 16 points higher) and two 2026 trials (Pakistan: doctors trained in AI literacy using GPT-4o scored 71.4% on diagnostic reasoning vs 42.6% with traditional resources, with GPT-4o alone at 82.9%; a UK replication found AI-assisted doctors still more than 21 points below the model alone). Physician Brandi Zachary says her team spent about a year, roughly $1 million and a 10-person clinical team and still has not finished building the system, because the hard part is not model intelligence but reliable data extraction, reproducible deterministic logic and safeguards. She warns a general-purpose model reading a lab report with 150 metrics may silently miss values — and an AI that does not know what it omitted is the dangerous kind.
It frames the real bottleneck for clinical AI as data extraction, reproducibility and human-machine division of labour rather than raw model reasoning.
多年来，布兰迪・扎卡里博士一直在把复杂的病史、化验结果、症状、用药和多年治疗经过，整理成连贯的临床方案。一个复杂病例，往往要花她三个小时研究，才敢确定什么重要、下一步怎么做。
后来，生成式AI出现了，能消化海量信息，调取医学知识的速度远超人脑，几秒钟就能看出一份病例里各项信息之间的关联。她以为，只要给大语言模型套个界面、加几道安全防护，就大功告成了。
实际上，他们花了一年左右、约100万美元，带着软件工程师和一支10人临床团队，系统却至今没建完。
她用的AI，和别人不一样
原因不是AI不够聪明，而是“聪明”恰恰是最容易的部分。
做临床软件之前她就发现：同一个AI工具，自己和旁边的人用，结果天差地别。因为她不是简单提问、拿走答案，而是在跟AI争辩、质疑假设、补上漏掉的背景、纠正错误、换角度再追问。