Microsoft Deserves Nobel Prize for AI Diagonosis Paper

I will start with an apology to the readers. My original title was “AI Will Not Replace Human Doctors”, but then I immediately realized that the Google bots would censor negative titles about this amazingly marketed technology. My conclusion is based on recent experience with two articles - “Alphafold is Amazing ! Here is a Tutorial” and “Alphafold, the Brawndo of Modern Science”. Both are critical of Alphafold, but one of them is censored heavily as I demonstrated here. Go figure !

Getting back to today’s topic, Microsoft researcher Harsha Nori recently published a paper titled Sequential diagnosis with language models. They showed that the AI “doctors” were better than the human doctors, and moreover do the job at a lower cost.

Using this benchmark, the system achieved 83.2% accuracy on New England Journal of Medicine clinicopathological conference cases and 80.3% on MIMIC-IV-ED emergency department admissions, outperforming generalist physicians (36.1 and 46.8%, respectively). Although these findings were obtained in a controlled benchmark setting, they highlight the promise of language models for supporting more accurate, lower-cost diagnoses.

AI boosters immediately jumped up and down on twitter pronouncing how most doctors would be replaced by computer programs. Some people from UK complained about Nigerian doctors taking over NHS, and argued that having a Microsoft program (actually GPT-5 from openAI) would be a welcome change. So, a chatbot with no nuanced understanding of human language, and which cannot tell that two articles are nearly identical because one is titled differently, will now take over the hospitals. I had to respond with “Nigerian Immigrant Radiologist at UK’s NHS Fixes Difficult Microsoft Bug Using AI”.

Here is a more serious response, and I like to highlight two issues with the primary claims of the paper being boosted by everyone.

1. Cost

The paper argues that the cost of chatbot “doctors” would be comparable with or lower than of human doctors, but the numbers on AI technology are not trustworthy. That is because this technology is being built with massive investments not counted in the estimates, and those investments are siphoned off from people’s pension funds and life insurances. The “price” that everyone sees on their openAI or Claude subscription is a heavily subsidized.

With budget constraints, DxO achieved 73.2% at a cost of $2634, and 69.5% at a cost of $2037. As expected, generalist physicians showed greater proficiency on ED cases than on CPC cases, with 46.8% (versus 36.1%) average accuracy. The average cost of testing by physicians on ED cases was $3400.

With all those investments properly accounted for, are chatbots 2x more expensive or 20x? The investors have to recoup their money, and so we may see a bait and switch with huge prices increases after people buy into this technology.

Such bait and switch is common in the tech world. Check Broadcom’s 10x increase of VMWare pricing for example. It is rather comforting to know Broadcom is a key company behind AI technologies.

2. Distrubution of Diagnosis Cases

The test bench used in the paper includes common and rare cases. I am not sure how much the performance numbers of chatbots are lifted by the rare cases.

Let me explain this point using an analogy. My scientific calculator has all kinds of buttons representing various functions, but the ones I use most frequently are the trigonometric and exponential ones. For the “rarely used” buttons, I will have to look up a manual. An equally weighted test on calculator usage will show that I am good with only “36.1%” of the buttons. However, I rarely encounter those remaining 63.9% of the buttons and that is why I am not proficient with them.

If a similar factor affects the benchmarks, then we are not getting an accurate representation of the difference.