Google's medical AI takes on Microsoft in hospitals

Search your symptoms online, and you usually end up convinced you have two days to live. Google knows the limits of basic web search all too well, which is why it is taking a much bigger swing: putting generative AI right inside hospital exam rooms.

According to The Wall Street Journal, Google has been testing its specialized medical language model, Med-PaLM 2, with healthcare partners including the Mayo Clinic since April. Unlike general chatbots, this system was trained on medical licensing exams to answer clinical questions, summarize dense documents, and organize sprawling health data.

Fluent answers, familiar hallucinations

The healthcare race between Big Tech is accelerating fast. Microsoft made its move in April by teaming up with health records giant Epic, integrating OpenAI models to draft patient responses. Both Google and Microsoft have pointed out that their models achieve comparable, high scores on licensing tests.

In fact, in research published by Google in May, physicians evaluating answers to over 1,000 medical questions actually preferred Med-PaLM 2's responses over human doctors across eight out of nine evaluation categories.

There is a massive catch, though. Those same reviewers found that Med-PaLM 2 generated more inaccurate or completely irrelevant information than actual doctors. Google researchers also admitted that the model showed no significant improvement in filtering out inaccurate statements compared to its initial version from December.

Even Google executives are urging caution. Greg Corrado, a senior research director who worked on the model, openly admitted: "I don't feel that this kind of technology is yet at a place where I would want it in my family's healthcare journey." Still, he argued that where AI helps, it expands efficiency tenfold.

To soothe ongoing privacy concerns, Google stressed that customer data remains encrypted, completely inaccessible to Google, and will not be ingested to train future models.

Why it matters

We think Google is walking a precarious tightrope here, trying to balance deep institutional caution against Microsoft's rapid enterprise land grab.

Teaming up with premier research hubs like the Mayo Clinic is a smart way to establish clinical credibility. But healthcare does not forgive hallucinations. A chatbot hallucinating a book citation is annoying; a clinical assistant confidently making up medical advice can be catastrophic. The research basically proves that LLMs are already exceptional at sounding authoritative, even when they are completely wrong.

For clinicians and health systems, Med-PaLM 2 is nowhere near ready for independent diagnostics. Its immediate value sits strictly in back-office grunt work: drafting routine patient messages, condensing hundred-page charts, and helping overburdened staff find data faster. Until tech giants solve the persistent hallucination problem, AI in medicine must remain an administrative assistant, not an intern with a stethoscope.