A developer named Antoine Finkelstein had had a sore right shoulder for weeks. After visiting an orthopedist, an MRI was performed on the affected area, and according to the medical report, he had a grade III partial tear in the subscapularis tendon. Finkelstein then did something else: he passed the resonance to Claude Opus 4.8 to see what the AI told him about that image, and the result was striking because according to that AI model his shoulder was “intact.”
AI gives you clues. The developer suspected that the clinic was perhaps trying to cash in on his condition, so he requested the raw DICOM data from the MRI. What they gave him was 266 MB which he crossed with currently available AI models. First, of course, he made a quick consultation with ChatGPT and in it he detected potentially important negligence: the clinic had applied shock wave therapy, which is not recommended for tendinopathies without calcification. He had also been injected with Traumeel, a homeopathic product registered in Germany “without therapeutic indication.”
Let’s see what Claude Opus 4.8 says. To try to get to the bottom of the matter, the user decided to turn the Anthropic model into a doctor to ask for a second opinion. After setting up the model on the Claude Code platform, he allowed the system to install the code packages needed to process the raw medical images he had been sent. After an hour of processing these images, the AI model issued a surprising diagnosis: the tendon that human doctors detected as 50% torn was completely intact.
I don’t trust. The result was so contradictory to the human diagnosis that Finkelstein wanted to go a little further and set up a blind arbitration system. He instructed Claude to deploy several independent subagents, combining AI images isolated from each other to avoid confirmation bias. The verdict of all those subagents was unanimous: there was no partial or total break, and everything suggested that the human specialists had exaggerated the diagnosis.
But quantity is not quality. This article gave rise to an interesting debate on Hacker News in which some important reflections were raised. It is important to remember, for example, that although AI eliminates the cost of consultations, having more information is not equivalent to solving the problem. As I said For one user, the situation reminded him of a problem he had with his car. He asked three different workshops for a solution, and each one told him something, and one even recommended a repair that he knew was useless. “The solution to uncertain information is not more information, which is certainly what AI can providebut better information, and right now AI can’t provide that.”
AI is too nice. There’s another problem here: big language models are meant to be nice and “nice.” They are in a sense echo chambers that want to keep us happy, so they are not designed to contradict us in a harsh way, which makes it easier to confirmation bias. If a user enters his suspicions in the prompt when asking the chatbot, the AI tends to agree with him: we often see how he begins by answering with “You are absolutely right…”. The problem with answers to medical topics is that they can be very different in independent sessions, but since the tone is always convincing and confident, they can lead to more confusion than the initial one.
The expert thinks. A professional radiologist participated in that conversation and provided expert insight. According to your criteriacurrent AI models remain mediocre at interpreting medical images due to the lack of massive public training databases. These data are protected by medical privacy laws, and at the moment this problem has a difficult solution, but that user explained that the latest models are already close in accuracy to that of a first or second year resident doctor. The theoretical threat to the radiology profession from AI is something we we have literally been talking for years: for now It doesn’t seem like something like this is close to happening..
Who is responsible. There is another big problem with AI: there is no one responsible if something goes wrong after applying a recommendation. It is true that human doctors can make mistakes and may have biases or even commercial incentives (selling treatments). However, the legal difference is fundamental: the medical system has a series of licenses, regulations and responsibility management that penalizes negligence. AI forces you to manage yourself in the face of uncertainty.
The problem is simple: trust AI or not. In issues as delicate as this, it is proven that AI is still far from being a real substitute for human experts. Today’s medicine may be “commoditized,” but AI, no matter how cheap or attractive it may seem, does not yet have the precision that would be needed for certain areas. As Finkelstein himself concluded, “I can’t know if I can trust the AI, so I’m in a kind of limbo in which I either try my luck with another doctor, or I wait and see if my shoulder improves with the rehabilitation I’m doing.”
Image | Vitaly Gariev
GIPHY App Key not set. Please check settings