AI Health Tech Advances, But Cures Lag Behind
▼ Summary
– AI has accelerated early-stage drug discovery, such as by computationally screening millions of compounds, but has not yet produced an FDA-approved drug.
– The clinical failure rate for drug candidates remains around 90%, as AI speeds up the process without demonstrably improving the odds of success in trials.
– AI health chatbots are ranked as a top technology hazard due to providing incorrect diagnoses and unvalidated medical advice to millions of daily users.
– A controlled study found that when real people use AI for symptom assessment, performance collapses and is no better than using traditional web searches.
– The core limitation is that AI cannot bypass the complexity of human biology or replace the essential conversational role of a physician in clinical care.
The promise of a healthcare revolution driven by artificial intelligence is everywhere, yet the reality for patients remains one of incremental progress and persistent challenges. While AI in drug discovery accelerates early research, it has not yet delivered a single approved therapy. Meanwhile, the widespread use of AI health chatbots poses documented risks, with safety experts ranking their misuse as a top hazard. The core issue is a widening gap between computational potential and tangible clinical outcomes.
Consider a recent project at Novartis. Researchers used generative AI to design 15 million potential compounds targeting Huntington’s disease. This computational feat narrowed the field to about 60 molecules synthesized for lab testing, yielding one promising scaffold. This process demonstrates extraordinary computational triage, yet it is not a cure. It highlights a critical distinction: AI excels at rapidly exploring vast chemical spaces, but translating a molecular scaffold into an effective, approved treatment for patients is a far longer and more uncertain journey.
The economic argument for AI in pharmaceuticals is compelling. Traditional development can take over a decade and cost billions, with a 90 per cent clinical failure rate. AI platforms can compress early discovery timelines dramatically. Insilico Medicine, for example, took an AI-discovered drug for idiopathic pulmonary fibrosis from concept to Phase II trials in under 30 months, a process that traditionally takes six to eight years. By early 2024, over 75 AI-discovered drug candidates had entered clinical trials. These are significant achievements in preclinical acceleration.
However, these achievements stop well short of the finish line. As of late 2025, no AI-originated drug has received FDA approval. The industry’s daunting failure rate in late-stage trials remains unchanged. Scientific analysis suggests AI-discovered compounds progress at rates similar to traditional ones, meaning the technology gets candidates to the starting gate faster without necessarily improving their odds of success. As one industry CEO noted bluntly, the last decade has seen “failure after failure” in delivering on AI’s therapeutic promise.
The fundamental limitation is not processing power but biological complexity. Diseases like Alzheimer’s or pancreatic cancer persist not because we screen molecules too slowly, but because their underlying mechanisms are poorly understood. AI cannot bypass the need for long and rigorous clinical trials that unfold over years in living human bodies. Novartis acknowledged this plainly at the World Economic Forum, stating AI is a tool for navigating complexity, not a magic wand. This is a defensible, if more modest, claim than visions of simply asking a chatbot to cure cancer.
If AI’s role in drug discovery is one of overstated progress, its deployment as a health assistant is increasingly a cautionary tale. In early 2026, the patient safety organization ECRI ranked chatbot misuse as the number one health technology hazard. These tools are not regulated medical devices, yet millions rely on them. ECRI documented instances of incorrect diagnoses, recommendations for unnecessary tests, and even the invention of a fictitious body part.
A landmark study in Nature Medicine underscored the problem. When tested in isolation, large language models correctly identified conditions in most cases. But when real people used them to evaluate their own symptoms, performance plummeted. Participants identified relevant conditions less than 35 percent of the time and chose correct next steps under 45 percent of the time, performing no better than a control group using standard web searches. The lead researcher, Dr. Rebecca Payne, stated clearly that AI is not ready to replace a physician. Medicine is a guided conversation, not a simple query, and chatbots cannot replicate a doctor’s ability to probe and clarify.
The situation in mental health applications is particularly concerning. The American Psychological Association has warned that chatbots and wellness apps are being used for purposes they were not designed for, such as treating psychological disorders. Research from Stanford found that therapy chatbots exhibited measurable stigma toward conditions like schizophrenia, a problem that did not improve with more advanced models.
This does not render AI useless in healthcare. AI-powered imaging tools are enhancing early cancer detection. Administrative applications that transcribe notes or summarize records are saving clinicians valuable time. These are genuine contributions, but they fall into a category of assistance, not autonomous intelligence. Dr. Payne offered a precise framework: the proper role for these models is as “secretary, not physician.”
The urgent health challenges remain. Alzheimer’s is projected to affect tens of millions, and survival rates for cancers like pancreatic have stagnated. These are the diseases where AI was heralded as a breakthrough hope. Yet, years into the generative AI era, its most visible health impact is tens of millions of people daily querying a chatbot about symptoms, while safety organizations urge extreme caution about the answers they receive. The technology is advancing, but the cures continue to lag behind.
(Source: The Next Web)
