A London-based dermatologist published results that should be uncomfortable for the artificial intelligence industry: a 2015 image model, fine-tuned for about twenty minutes, outperformed two of the world's most powerful language models at diagnosing skin cancer.
Dr. Magnus Lynch, a Mohs surgeon at Guy's Hospital and honorary researcher at King's College London, tested GPT-6 Astra and Claude Fable 5.1 — the most capable models available from OpenAI and Anthropic in October 2026 — against an Inception v3 network specifically fine-tuned for skin lesion recognition. The result was clear: the specialized model achieved 62% accuracy on first-choice diagnoses, compared to 42% for both LLMs. And for benign lesions, the disparity was even wider.
The test was straightforward but revealing. Lynch selected 200 images of biopsy-confirmed lesions from the public MILK10k dataset, available since 2025. For each case, he presented two images to each model: a standard clinical photograph and a dermoscopic image (skin microscopy). He then asked each model to classify the lesion among six possible diagnoses.
The LLMs failed in systematic and predictable ways. GPT-6 Astra labeled 75 lesions as basal cell carcinoma when only 25 were — including 14 of the 30 melanomas in the set. In other words, instead of identifying melanoma, the model confused it with basal cell carcinoma, a much less aggressive condition. Claude Fable 5.1 showed a different bias: it labeled 53 lesions as actinic keratosis against only 20 real cases. Both models demonstrated a consistent over-prediction bias, tending to label any ambiguous lesion as their preferred diagnosis.
The Inception v3, though old and specialized, did not make that mistake. Of the 30 melanoma lesions in the test, the model correctly identified 13 as its first choice. GPT-6 Astra got 8 and Claude Fable 5.1 got 11. For benign lesions, the difference was dramatic: the CNN correctly recognized 37 of 50 moles and 26 of 50 seborrheic keratoses, compared to 16 and 13 for GPT-6 and only 33 and 6 for Claude.
What makes this result particularly interesting is not the performance of the Inception v3 itself — it is a baseline model with an AUC of 0.86 on this dataset, well below the 0.96 achieved by modern networks fine-tuned for dermatology. What matters is what it reveals about generalist LLMs.
Large language models have been trained on texts, code, medical articles, and clinical manuals. They know about dermatology. But knowing the theory is not the same as recognizing visual patterns in real images. The Inception v3 knows nothing about melanoma beyond its convolutional layer architecture. It was exposed to only images and labels, fine-tuned for twenty minutes on a specific problem, and on that narrow task outperformed the world's most expensive generalist models.
This does not mean LLMs are bad at medicine. It means that language generalization does not directly translate into diagnostic visual competence. LLMs are, essentially, statistical models operating in text space — they learn that "dark irregular border" often correlates with melanoma in medical texts, but they do not see the irregular border. They see tokens.
Lynch listed reasonable limitations of the study. Two hundred cases are sufficient to show the pattern, but not to rank close results. Each LLM run was made only once, although LLM responses vary between runs. And MILK10k has been public since 2025, so it is possible that both GPT and Claude have seen these images in their training data.
But even with all caveats, the pattern holds: the world's most advanced models failed at a basic medical visual task for which an old, narrow model performed significantly better.
The implication is not that the Inception v3 will replace dermatologists. It is that the artificial intelligence industry, by betting almost entirely on generalist language models, may be neglecting the importance of specialized models for critical visual tasks. In a market where digital health is one of the fastest-growing sectors, specialization still beats generalization — and this applies to diagnostics, medical imaging, and likely much more that we are only beginning to explore.
Sources: Lynch Lab, AI Chat Daily, AI Governance
✓ Independent sources cross-checked and verified before publishing
