Can you outsource your decision making around FHIR to AI?
Can you rely on Claude or other generative AIs to do the job your FHIR people are doing today?
I put that to the test last week and wrote a detailed report about it for my mailing list.
You can read that report here.
I wanted to see if Claude could solve what’s becoming a common data capture request in FHIR: How to identify Observation resources generated by an AI of some kind.
Claude failed to come up with a good way of doing this.
It got there in the end, but only after I repeatedly nudged it in very specific directions and repeatedly questioned its poor choices.
What I found particularly unsettling was the confidence with which Claude held onto and backed up its poor selections. If I hadn’t known the answer in advance I might have been convinced to accept its recommendations, leading to a poor outcome.
Once it got to the “correct” solution (there are more than one) it provided excellent and articulate reasons for doing so. There’s value in that from a documentation perspective, but it needed far too much guidance to get there.
The lesson here is that AIs like Claude can be useful when working with FHIR, but only if you have the knowledge to know when it’s going wrong or off-track.
Useful in experienced hands but dangerous for beginners.
Here’s the final artifact produced by Claude: https://darrendevitt.com/wp-content/uploads/2025/08/FHIR-AI-Attribution-Artifact.pdf
Spot the glaring mistake on the last page of the document!
And here’s the published series of prompts and responses: https://claude.ai/share/633cb422-5614-4f15-91ca-f65dff093270
---