I used Claude Code to get a second opinion on my MRI
Using Claude and other large language models to get a “second opinion” on MRI scans and other medical images is generating both curiosity and alarm. Commenters note that current frontier models can be useful for explaining text reports and suggesting questions to ask a doctor, but are poorly validated for image-based diagnosis, prone to confident hallucinations, and may not actually use the images provided. Many see AI as a promising adjunct for informed patients and overburdened health systems, yet warn that relying on it for radiology or complex care risks misdiagnosis, erodes trust in clinicians, and can be especially dangerous when people already struggle to access reliable medical expertise.
Limits of LLMs for Medical Imaging
- Many commenters strongly argue current general-purpose LLMs are not reliable for image-based diagnosis, especially complex 3D modalities like MRI.
- Radiology and MRI researchers note frontier models are not trained or validated for subtle medical image interpretation and tend to hallucinate plausible-sounding findings.
- Several references to work showing “mirage reasoning”: models confidently describe and reason about images that were never actually provided.
- Practitioners report benchmarking frontier models on real medical image datasets (e.g., otoscopy, pathology slides) and finding poor calibration and high-confidence errors.
Radiology & Imaging Nuance
- MRI details: common “2D” slice protocols with gaps vs true 3D isotropic scans; tradeoffs between resolution, speed, and motion artifacts.
- Ultrasound in orthopedics is useful for soft tissue and tendons but poor for bone and small calcifications; plain radiographs or MRI can detect calcifications that ultrasound misses.
- Radiologists stress that negative findings are always conditional on modality and technique; “no X” on ultrasound doesn’t mean “no X” absolutely.
AI as Second Opinion / Patient Tool
- Some users share positive anecdotes where LLMs helped spot misdiagnoses or suggest overlooked conditions/treatments, especially from text reports and lab data.
- Others report blatantly wrong or internally inconsistent results on their own MRIs/X-rays, or obvious failures like misreading chessboards and basic images.
- A recurring “best use” pattern: use LLMs to understand reports, generate questions, or surface guidelines and differential diagnoses, then discuss with a human clinician.
Trust, Self‑Diagnosis, and “AI Psychosis”
- Clinicians worry about patients treating AI outputs as authoritative, eroding trust and consuming scarce visit time to debunk confident nonsense.
- Multiple comments liken this to an amplified “Dr. Google” problem, with extra danger because the answers sound expert and personalized.
- Some describe a quasi-religious overconfidence in AI (“AI psychosis”), where contrary expert feedback is dismissed as lack of vision.
Healthcare System Problems & Incentives
- Many stories of misdiagnosis, conflicting opinions, overtreatment (e.g., unnecessary imaging, procedures, homeopathic injections), undertreatment, and rushed 5–15 minute visits.
- This drives patients toward LLMs out of frustration and lack of access, not just techno-optimism.
- Several argue AI could be powerful as an internal tool for doctors (guideline lookup, lead generation, checklists), but not as a stand‑alone diagnostician, especially for imaging.