Medical AI is worse than Google for diagnosing your symptoms (scientific proof)

Imagine typing your symptoms into an AI chatbot and hoping for a reliable diagnosis, as in a sci-fi movie. However, a recent British study reveals that these tools do not even outperform a simple Google search when real patients turn to them. With 40 million daily users consulting ChatGPT for health advice, the enthusiasm is growing, but the limitations are clear.

AI powers pharmaceutical research without valid miracles

By the end of 2025, scientists at Novartis used generative AI to generate 15 million potential compounds against Huntington’s disease, degraders capable of crossing the blood-brain barrier. Only 60 have been produced in the lab, leading to a promising lead in the optimization phase. A leap in virtual screening, but far from a drug ready for use.

As of today, no AI-discovered treatment has received FDA approval. In January 2024, the Boston Consulting Group counted 75 candidates in clinical trials boosted by AI, shortening preclinical stages from several years to 13-18 months and overall timelines by 30-40%. The exorbitant costs – $2.5 billion per approved molecule, 90% clinical trial failures – persist. Expert Raminderpal Singh insists: the true measure is in phase III, justifying the industry’s entirely justified caution. An anonymous executive laments: “AI has disappointed us all in the past decade. We have only seen failure after failure.” Novartis at the World Economic Forum in January 2026 admits: “AI is not a magic wand. It is a tool to navigate complexity more intelligently.” Human biology remains a quickly unsolvable puzzle, despite billions of sorted molecules.

Chatbots: the number one danger for amateur diagnostics

In January 2026, the ECRI placed AI chatbots at the top of the list of health technological risks, with cases of false diagnoses, unnecessary tests, or absurd anatomical inventions. Not classified as medical devices, they proliferate without safeguards.

Researchers at Oxford tested this in February 2026 through a randomized study on 1,298 volunteers, published in Nature Medicine. AI alone identified 94.9% of pathologies; when assisting patients, the score dropped to 34.5%, equivalent to a basic web search. Rebecca Payne, an Oxford physician, concludes: “Despite all the hype, AI is simply not ready to replace the doctor.” Humans struggle to accurately describe their ailments, unlike professionals who delve into targeted questions. In mental health, the American Psychological Association warns: these tools are not equipped for it, and Stanford demonstrates biases against conditions like schizophrenia.

AI truly helps, far from the role of a virtual doctor

AI excels in image analysis to quickly track certain cancers or in administrative tasks to free up caregivers. It accelerates the early stages of drug hunting but struggles with nuanced clinical reasoning. Rebecca Payne succinctly sums it up: a role of “secretary, not doctor.”

Challenges are mounting: 78 million affected by Alzheimer’s in 2030, a five-year survival rate of 13% for pancreatic cancer despite general progress reaching 70% for all cancers. Three years after the explosion of generative AI, millions still turn to these tools every day for a simple headache.

In our geek era where AI promises foolproof diagnoses like JARVIS, the reality is: powerful tools as supplements, never as soloists. It is better to have an appointment with the doctor than to rely on a poorly made quick fix.