Claude Code read my MRI and contradicted the doctor's torn-tendon diagnosis

A developer with weeks of shoulder pain got an MRI at a clinic that diagnosed a Grade III partial-thickness tear of the subscapularis tendon and immediately began an aggressive, repeat-treatment plan. Suspicious of how fast they moved — and after GPT-flagged two questionable interventions (shockwave therapy not recommended for non-calcified rotator-cuff tendinopathy, plus an injection of a homeopathic product with no therapeutic indication) — he fed the raw 266 MB DICOM export to Opus 4.8 running inside Claude Code, where the model could install packages and run real analysis code rather than just chat.

After about an hour of planned work, Opus reported the tendon as intact, directly contradicting the radiologist’s reading. To adjudicate, he had Claude compare both reports using fresh subagents to avoid bias; the ‘arbiter’ sided with the AI reading at moderate-to-high confidence, finding only mild insertional tendinosis and no discrete tear. Notably, the AI received less clinical context than the human doctors did.

The author is careful to frame this as technical curiosity, not medical advice — he isn’t a doctor, the AI could be wrong, and he may have misread the situation. The real takeaway is the unsettling middle ground AI now creates: enough credibility to undermine trust in a costly, intervention-heavy diagnosis, but not enough to fully trust instead. He’s left choosing between a second human opinion and waiting out rehab, hoping future models earn the trust we already give AI for mundane tasks.