On Monday, a Finnish developer named Antoine published a blog post titled “Using Opus 4.8 to get a second opinion on an MRI and where it leaves me.” He’d fed 266 megabytes of DICOM imaging data from a shoulder scan into Claude Code, along with the radiologist’s report and a prior conversation with ChatGPT 5.5 Pro. The model disagreed with the human diagnosis. By Tuesday morning, the post had 360 points on Hacker News and over 470 comments, the comments section doing what HN comments sections do: half earnest debate about pixel-level analysis, half hand-wringing about whether this is responsible.

I want to bracket the question of whether Antoine should have done this. The disclaimer-heavy, deeply hedged tone of the post suggests he’s already absorbed that lecture. What interests me is what the post inadvertently documents about the structure of medical uncertainty — and who gets to distribute it.

The Asymmetry Nobody Mentions

Antoine’s MRI came with a human radiologist’s report and a suggested course of treatment that, in his telling, was extensive enough to begin within minutes of the scan. He had a diagnosis. He had a treatment plan. He had access to the medical system. And he still went looking for a second opinion from a language model.

This is not the story of a desperate patient with no options. It’s the story of a well-resourced patient with options who decided they weren’t enough. That distinction matters, because the policy conversation around AI in medicine almost always runs in the opposite direction: AI will democratize access, bring specialist-level diagnosis to underserved populations, close the gap between the Mayo Clinic and a rural clinic in the developing world. That’s a noble vision. It’s also not what happened here.

What happened here was a knowledge worker with a browser tab open to Claude Code deciding that the existing system — which includes a radiologist, a referring physician, and presumably a national healthcare apparatus — was insufficiently thorough. Not because it was unavailable, but because it was unsatisfying. The model didn’t fill a gap in access. It filled a gap in confidence.

The Hotel Bar Problem

Second opinions are not new. What’s new is the friction. A traditional second opinion requires scheduling another appointment, traveling to another clinic, waiting for another radiologist to review the same images, and then sitting in another examination room while someone explains whether the first person was right. That process costs time, money, and social effort. It also imposes a natural ceiling on how many opinions a patient can reasonably seek.

Claude Code removes that ceiling. Antoine didn’t just get a second opinion; he got one after feeding the model his ChatGPT conversation, effectively creating a third opinion that had read the first two. This is not a bug of the system. It’s the feature. And it raises a question the medical establishment is not ready to answer: if a patient can get infinite second opinions for the marginal cost of an API call, what does the doctor’s authority actually rest on?

One radiologist I spoke with — in a hotel bar at RSNA last year, not on the record — put it bluntly: “The report is a negotiation. It always has been. The difference is the patient used to not be in the room.” What he meant was that radiology reports have always contained uncertainty. Phrases like “cannot exclude” and “correlate clinically” are not failures of precision; they’re acknowledgments that imaging is interpretive. The model doesn’t eliminate that interpretation. It just makes it visible to the patient in a way it wasn’t before.

What the HN Thread Missed

The Hacker News discussion, predictably, got bogged down in questions of accuracy. Can Claude read DICOM files? Did it hallucinate? Should Anthropic have guardrails that prevent this kind of use? These are important questions. They are also the wrong ones.

The more interesting question is structural: the blog post is a signal that the bottleneck in medicine is shifting. For decades, the bottleneck was information. Only a radiologist could look at an MRI and tell you what it meant, because only a radiologist had the training to interpret the images. That bottleneck is dissolving. What’s replacing it is not a new bottleneck of information — it’s a bottleneck of trust.

Patients don’t lack data. They lack a way to adjudicate between competing interpretations of data. Antoine now has two opinions: his doctor’s and Claude’s. They disagree. He has no way to resolve the disagreement except by seeking a third opinion, which will likely disagree with at least one of the first two. This is not a failure of AI. It’s a failure of the institutional frameworks that were supposed to make medical authority legible.

The Uncomfortable Thing We’re Actually Building

The medical AI revolution isn’t going to look like a rural clinic in sub-Saharan Africa getting a superhuman diagnostician. It’s going to look like Antoine’s blog post: a well-insured, well-educated patient in a wealthy country using a frontier model to double-check a diagnosis he already received from a board-certified specialist. The technology will not democratize medicine so much as it will arm the already-empowered with more questions to ask.

That’s not an argument against the technology. It’s an argument against the story we tell about it. If the primary use case for medical AI is affluent patients auditing their own diagnoses, then the policy question is not “how do we ensure equitable access to AI diagnostics?” It’s “how do we redesign the physician-patient relationship when the patient arrives with a model’s dissent in hand?”

Nobody in the 470-comment HN thread had a good answer to that. Neither do I. But Antoine’s shoulder MRI, whatever it actually shows, has made one thing clear: the second opinion is no longer scarce. What’s scarce is the authority to say which one counts.

Sources