Doximity Outranks OpenEvidence, Frontier Models in Independent Stanford-Harvard Study of Clinical AI Safety - Financial Times

Ah, Stanford and Harvard have finally answered the burning question: which AI can give you a diagnosis without also inventing a cure for gravity? Their NOHARM benchmark — the *first* independent clinical safety study, naturally — crowns Doximity Ask as the mightiest medical chatbot. Beating frontier models? A thumping victory for a niche tool that probably doesn't try to write poetry about your MRI.

Let's savour the irony: the 'leading frontier AI models' — the ones racking up billions in compute and promising AGI — apparently can't keep a clinical conversation safe. Doximity, a company you've never heard of unless you're a doctor who needed a fax machine, tops the chart. It's like discovering that a plumber fixes pipes better than a PhD in hydraulic engineering. The frontier labs are busy chasing omni-modal superintelligence; meanwhile, a specialised medical scribe shows that maybe, just maybe, a narrow focus beats a jack-of-all-trades.

The study, led by ARISE (because every benchmark needs a slightly pompous acronym), looked at real-world clinical samples. That means actual patients. Actual risks. And Doximity didn't just win — it 'outperformed leading frontier AI models.' We suspect the frontier models are too busy hallucinating a new specialty called 'imaginary oncology.' So yes, the safest clinical AI is the one nobody's heard of. That's not reassuring. That's a warning flare about the emperor's new neural network. But hey, at least the fax machine industry has a future.