Dr Magnus Lynch, a consultant dermatologist at Guy’s Hospital in London, gave 200 biopsied skin lesions to three AI systems and counted how often each named the right diagnosis. A network Google published in 2015 won, ahead of both chatbots. Lynch Lab posted the results on 25 October 2026.

The old system is Inception v3, which belongs to a family called CNNs, short for convolutional neural networks: image-recognition software that learns to pick out patterns in pixels and does nothing else. Lynch fine-tuned it on 68,011 skin images from a public archive, about twenty minutes of work on one H100 chip. Against it he set GPT-6 Astra and also Claude Fable 5.1, which he calls two of the most capable models on the market this month.

Every lesion came as two photos: an ordinary close-up plus a dermoscopic image, shot through the lit magnifier dermatologists use. Half the set needed follow-up: 30 melanomas, 20 actinic keratoses, and 50 other skin cancers split evenly between basal cell and squamous cell types. The other half were 50 moles and 50 seborrhoeic keratoses, which are harmless warty growths. All 200 had been biopsied, so they skew hard. Nobody biopsies a mole unless it bothered a clinician first.

Given six possible answers, the CNN got 125 right, or 62%. GPT-6 Astra got 84 and Claude Fable 5.1 got 85, both 42%. Allowing a second guess, the CNN reached 173, against 134 and 138. On melanoma it named 13 of 30, while the chatbots managed 8 and 11.

The widest gap came on harmless lesions. The CNN recognised 37 moles out of 50, and 26 seborrhoeic keratoses out of 50. GPT-6 Astra managed 16 and 13, and Claude Fable 5.1 managed 33 and 6. The chatbots did win a few rounds: GPT-6 Astra identified all 25 basal cell carcinomas to the CNN’s 24, and both chatbots beat it on actinic keratosis. Each chatbot also leaned on one pet answer. GPT-6 Astra labelled 75 lesions as basal cell carcinoma when only 25 were, 14 of the 30 melanomas among them.

The part a clinician would care about most is the discharge question. Lynch asked each system whether a lesion could simply be sent home, with nothing further done. Of the 30 melanomas, GPT-6 Astra would have sent 3 home, and Claude Fable 5.1 would have sent 4. The CNN, at the threshold it was set to before testing, would have discharged 18, and it kept only 77 of the 100 lesions that needed attention. It also sent home 76 harmless lesions, against 32 and 22 for the chatbots. When Lynch tuned its threshold to match the chatbots on unsafe lesions, it discharged one or two melanomas. So the CNN’s edge is in naming the diagnosis. As a gatekeeper it depends on where someone sets the dial.

Lynch lists his own limits. Two hundred lesions cannot separate close results, and with only 30 melanomas the most important number is shaky. Each chatbot case ran once, and answers vary between runs. The images went public in 2025, so the chatbots may have met them in training. He also calls Inception v3 a baseline: regulated skin-cancer products built on newer systems, he writes, perform considerably better. In a separate test, he showed each chatbot only the clinical photo and asked as a patient would, “do I need to see a doctor?” Both advised a visit in almost every case, including 99 and 98 cases out of another 100 melanomas confirmed by biopsy. He also discloses that Anthropic’s coding tool Claude Code produced the analysis and first drafts, and that one of the tested chatbots is Anthropic’s.

None of this says a chatbot should read your mole, or that a 2015 network should. It does say that a cheap, narrow model trained on the right images can beat a general one on a narrow task, which is a useful thing to know before paying for the bigger model. Lynch published the cases, prompts, every chatbot reply and the scoring code, so any team weighing a chatbot for image triage can rerun the 200 cases against its own setup before buying.

Reported by Lynch Lab (Dr Magnus Lynch) on 25 October 2026.