🧑 You
Why was this thread so controversial: https://news.ycombinator.com/item?id=48708941
▶ Tool: fetch_url [url='https://news.ycombinator.com/item?id=48708941']
▶ Tool: fetch_url [url='https://antoine.fi/posts/mri/']
▶ Tool: web_search [query='antoine.fi "Claude Code" "second opinion" MRI']
▶ Tool: fetch_url [url='https://antoine.fi/mri-analysis-using-claude-code-opus']
🤖 Assistant
The post "I used Claude Code to get a second opinion on my MRI" (by antoine.fi) sparked a highly contentious 600+ comment discussion. Here's what made it so inflammatory:
---
The author uploaded his raw shoulder MRI (DICOM files, ~266 MB) to Claude Code (Opus 4.8) and asked it to analyze the images. Claude concluded his tendon was intact — directly contradicting the human radiologist's finding of a Grade III partial-thickness tear. An arbitration phase also sided with the AI.
To many HN readers, this was deeply irresponsible — using a general-purpose LLM (not a specialized medical imaging model) as a diagnostic tool. A practicing radiologist commented (link):
> "I can't conclusively disagree with the original report... These models are generally terrible at reading medical images... Worse than useless."
The comments split into two heated factions:
|
Anti-LLM Diagnosis Camp |
Pro-LLM/Second Opinion Camp |
|
LLMs produce plausible but not factual outputs |
Human doctors misdiagnose all the time |
|
Non-experts cannot verify AI claims |
Many shared stories of harmful misdiagnoses |
|
One commenter: "Ask 5 LLMs, get 5 contradictory answers" |
Author's doctor pushed shockwave therapy (against guidelines) & homeopathic injections |
|
"Absence of evidence ≠ evidence of absence" |
Medical system is opaque, expensive, and hard to navigate |
|
A radiologist called it "worse than useless" |
AI can be a tool for patient advocacy and informed questioning |
1. The Babbage quote fight: Someone invoked Charles Babbage's famous quote about putting wrong numbers into a machine — sparking a multi-level meta-argument about whether LLMs are "thinking" or "bullshitting."
2. The mechanics analogy: Multiple car repair anecdotes debating whether more information/opinions actually helps or just creates confusion.
3. The "you can't verify this" problem: The central trap — if you're not a radiologist, you can't tell if the AI or the human is right, so the AI "second opinion" gives a false sense of clarity.
4. The medical system critique: Several heartbreaking personal stories of misdiagnoses by human doctors, with one commenter having to fight being forcibly committed to a tuberculosis hospital due to a radiologist's error.
5. Prompt engineering as undefined behavior: One memorable quote: "I'm using this new programming language: it's called LLM prompting, and everything is undefined behavior."
The thread's controversy ultimately boils down to a question that has no good answer:
> If you can't fully trust the doctor, and you can't fully trust the AI... what do you do?
The author himself ended up in "limbo" — stuck between a human expert he didn't fully trust and an AI he knew could hallucinate. Many commenters found this existential uncertainty about modern medicine + AI deeply unsettling, while others found it empowering.
- 472 points on HN (highly upvoted)
- 606 comments (very high engagement)
- The thread stayed active for over 20 hours
- It touched on: radiology, epistemology, medical malpractice, AI safety, homeopathy, the placebo effect, and the nature of trust