Generative AI and diagnosis: an anecdote

10/4/26

This is not the first anecdote1 I’ve read about how AI can help support people with complex medical issues solve them, though it may be the most dramatic. Our AI Midwife is the tale of a couple struggling with infertility who discovered, after an MRI conducted and analysed by humans, that the issue was caused by a large, previously undetected, fibroid that happened to be right in prime implantation territory.

This isn’t as surprising as it might seem. Dr Google was problematic partly because the data returned wasn’t personalised. LLMs have injested a heck of a lot of information, if you hand them a bunch of data about what’s wrong and get it to analyse it (something computers are really good at) they’re statistically likely to produce the right answer occasionally.

LLMs have undeniably had a positive impact for this couple, and yet the author is still apologetic about using them. He is right: much “of the discourse around AI has become toxic” to the point where people using it thoughtfully feel they have to justify it.

This technology has a lot of problems but it’s not unambiguously bad. Like every other technology the outcome is has more to do with how it is used then any properties inherent in it.2

  1. the plural of anacdote isn’t data ↩︎

  2. I am aware that this is dangerously close to the gun lobby’s “guns don’t kill people, people kill people”. There is a debate to be had around access, ownership, guardrails etc. But we can’t have it if we refuse to accept that people are using it and that the outcomes aren’t always awful. ↩︎

AI