What Happened
A TIME investigation examined why AI has failed to deliver on its healthcare promises despite years of bold predictions. The numbers tell a complicated story: nearly 950 AI and machine learning tools received FDA approval between 1995 and 2024, with 723 of those being radiology devices. Yet radiologist employment has increased, not decreased, despite AI pioneer Geoffrey Hinton's 2016 prediction that AI would surpass them within five years.
The report highlights several critical failures. ChatGPT's advanced model triaged medical emergencies incorrectly more than 50% of the time. Stanford cardiologist Jack W. O'Sullivan found that AI systems produced clinically significant hallucinations at a rate of 6.5%, though they self-corrected when questioned. Meanwhile, at least 12 million diagnostic errors occur annually in the U.S., causing roughly 800,000 cases of disability or death - a problem AI was supposed to help solve.
Five separate studies showed that AI systems sometimes outperform physicians, but physicians using AI as a tool performed inconsistently. Researcher Eric Topol noted he is "not as confident as I was in 2019" about AI's medical trajectory.
Why It Matters
This matters beyond healthcare because it exposes the reliability ceiling that all AI tools face when stakes are high. The same hallucination and overconfidence problems that cause a 6.5% clinical error rate in medical AI also show up in coding assistants, legal tools, and business analytics - just with lower consequences.
The "automation neglect" finding is particularly relevant. Doctors anchor on their initial diagnosis and ignore AI suggestions. This is the same pattern we see with AI coding tools: developers accept the first suggestion and stop critically evaluating. In medicine, that kills people. In software, it ships bugs.
The legal asymmetry is also worth noting. Doctors face liability for using AI tools that give bad advice, but face no consequences for ignoring available AI tools that could have caught an error. This creates a rational incentive to avoid AI entirely, which slows adoption regardless of how good the technology gets.
Our Take
Healthcare is where the "good enough" standard that works for content generation and code assistance completely breaks down. A 93.5% accuracy rate sounds impressive until you realize it means 6.5 out of every 100 medical decisions include a significant hallucination.
The most honest quote in the piece comes from Hinton, who backtracked from his radiologist prediction by reframing: "Healthcare is a very elastic market... we'd just all get ten times as much healthcare." That is a polite way of saying AI will not replace doctors - it will generate more work for them.
The real opportunity the article identifies is upstream: AI analyzing sleep data from 500 million smartwatch users, predicting conditions from wearable sensors, shifting medicine from treatment to prevention. Stanford researchers demonstrated 130 conditions could be predicted from a single night of sleep sensor data. That is where AI's pattern-matching strengths align with healthcare needs without requiring the kind of reliability that direct diagnosis demands.
For anyone evaluating AI tools in high-stakes domains, the lesson is clear: AI works best as a screening layer and worst as a decision-maker. Build workflows accordingly.