Imagine you need someone to identify birds in your garden. You could hire the world's most well-read librarian — someone who has absorbed every book ever written, including field guides — or you could hire a birdwatcher who has spent ten years staring at nothing but birds through binoculars. The librarian knows what a robin is in theory. The birdwatcher knows what a robin looks like at dusk, from below, in molting season. This paper is about what happens when you give both of them the same 200 birds and ask them to name each one. Dr. Magnus Lynch, a consultant dermatologist at Guy's Hospital in London, tested GPT-6 Astra and Claude Fable 5.1 against a fine-tuned Inception v3 — a convolutional neural network architecture Google published in 2015 — on 200 biopsy-proven skin lesions from the public MILK10k dataset. Each lesion had both a clinical photograph and a dermoscopic image. All three systems chose from the same six diagnoses. The CNN got 125 right (62%). Both frontier LLMs got 84 and 85 right (42%). Allowing a second choice, the CNN reached 173 of 200 against 134 and 138. The test design is deliberately conservative in ways that matter. All 200 lesions were biopsied, meaning even the benign ones were clinically suspicious enough to warrant cutting — these are harder cases than average. The CNN was fine-tuned on 68,011 dermoscopic and clinical images from the ISIC Archive on a single H100 GPU, with every test image and every image of the same patient removed from training and validation data. The LLMs received the same prompt at default settings with both images and nothing else. Contamination, if anything, favored the LLMs: MILK10k has been public since 2025 and could have appeared in their training data. The failure modes are as telling as the accuracy numbers. GPT-6 Astra had a basal cell carcinoma fixation, labeling 75 lesions BCC when only 25 actually were — including 14 of 30 melanomas. Claude Fable 5.1 over-called actinic keratosis, tagging 53 lesions with that diagnosis against a true count of 20. The CNN's melanoma errors went the opposite direction: of 17 melanomas it missed, it called 8 a naevus and 6 a seborrhoeic keratosis. In clinical terms, the LLMs' biases are more dangerous — calling a melanoma a BCC or AK at least keeps the patient in the system, but both LLMs also discharged 3 and 4 melanomas respectively at their default output. The CNN's safety profile introduces a necessary nuance. At its preset threshold, it discharged 18 of 30 melanomas — far worse than the LLMs. But when thresholds were matched to the LLMs' keep rates (keeping 97-99 of 100 unsafe lesions), the CNN discharged only 1-2 melanomas while correctly releasing 76 safe lesions versus 22-32 for the LLMs. This is the operating-point problem: the CNN can be tuned for clinical conservatism, and when it is, it dominates on specificity without sacrificing sensitivity. The structural lesson is not that LLMs are bad at medicine. It is that general-purpose language models, even frontier ones, are not vision specialists. Inception v3 is explicitly described as not state of the art — modern skin cancer AI uses ensembles of vision transformers trained on larger datasets, and regulated products built on them perform considerably better. The CNN's 62% is a floor for what domain-specific fine-tuning achieves, not a ceiling. The LLMs' 42% is what you get when you ask a system optimized for language reasoning to do pattern recognition on dermoscopic images. Lynch disclosed that analysis and first drafts were produced with Claude Code (an Anthropic tool), and one of the two tested LLMs is Anthropic's Claude Fable 5.1. Scoring was done by script against histopathology, removing subjective judgment from the evaluation. The CNN runs entirely in a phone browser via WebGPU — nothing is uploaded — and the full dataset, prompts, every LLM reply, and scoring code are publicly available.
