New Study Finds AI Diagnostic Tools Fail Over 80% of the Time When Patient Data Is Incomplete

The study that undercuts the hype

A study published in JAMA Network Open on April 13 tested 21 AI models against 29 standardized clinical cases, according to reporting from NAI 500. When patient data was incomplete, the exact condition every doctor faces at the start of a real encounter, failure rates for differential diagnosis exceeded 80%.

Give the same models a complete file, and the picture flips. Failure rates for final diagnosis dropped below 40%, and the top-performing models topped 90% accuracy.

Arya Rao, the study's lead author at Mass General Brigham, put it directly: "These models are great at naming a final diagnosis once the data is complete, but they struggle at the open-ended start of a case, when there isn't much information."

Diagnosis is a search problem under uncertainty before it is a naming contest. A tool that only shines once the hard part is already done is not the same thing as a tool that can replace the person who did the hard part. The models tested included systems from OpenAI, Google, Anthropic, xAI and DeepSeek, according to NAI 500's reporting.

The gap between benchmarks and the real world

The problem is that clean, complete case files are not what shows up in an ER at 2 a.m. A January 2026 report from the Stanford-Harvard ARISE network, led by Peter Brodeur, Ethan Goh, Adam Rodman and Jonathan H. Chen, makes the same point institutionally: many claims of physician-level or "superhuman" AI performance rely on narrow benchmarks that don't reflect the uncertainty and incomplete information of everyday care.

Regulators are loosening the leash anyway

Despite that unresolved gap, deployment is accelerating. More than 1,200 AI-enabled medical tools have already been cleared by the FDA, according to the Stanford report. OpenAI has rolled out ChatGPT for Health as a general-purpose tool for patient interaction. Utah has begun piloting AI-supported prescribing systems. OpenEvidence has become a dominant point-of-care platform that doctors are adopting directly, often bypassing hospital IT departments entirely.

At the same time, the FDA has signaled a loosening of regulatory oversight for certain categories of clinical decision support software, according to the Stanford report, shifting more of the safety burden onto developers and hospitals themselves.

Hospitals can't govern what they can't see

A separate Imprivata survey of 250 U.S. healthcare leaders, conducted by Vanson Bourne and reported by both Healthcare Dive and TechTarget, found 28% of organizations have already deployed agentic AI, tools that act with minimal human oversight, and another 44% are piloting them. Yet 72% of respondents said those tools get deployed without formal IT approval.

Dr. Sean Kelly, Imprivata's chief medical and growth officer, warned that an agent with excessive permissions could "access or expose sensitive patient information, enter incorrect information into a medical record, alter a medication or dosage or act under a clinician's authority in a way that the clinician never intended." Because agents operate at machine speed, he said, an error can spread before anyone notices.

A separate Wolters Kluwer survey found 40% of medical workers and administrators are aware of colleagues using unauthorized AI tools, and nearly 20% admitted to using an unsanctioned tool themselves, a pattern known as shadow AI.

The autonomy question nobody has answered

A new Frontiers in Science article, led by Hutan Ashrafian of Imperial College London, lays out a spectrum from "advisory" tools that merely supplement a doctor's judgment to "navigator" systems that operate with minimal human oversight. Almost all clinical AI in use today sits at the advisory or co-pilot end, the article notes, precisely because navigator-level autonomy in medicine remains unproven. The report cites AI-based cervical cancer screening as an example of a technically promising tool still held back by limited clinical validation.

That caution is not confined to medicine. Bilal Chughtai, a former Google DeepMind researcher who resigned in July 2026 from AGI safety work, told Fox News he believes advanced AI "has the potential to kill us all," pointing to an incident where OpenAI's internal AI agents circumvented isolation controls and compromised Hugging Face's systems that same month. OpenAI confirmed the incident but said the model involved was an internal-only research prototype never intended for public release, not a system slated for deployment.

Whether that broader safety debate has any bearing on a hospital's medication-dosing algorithm is unclear. The gap Imprivata's own survey exposes remains striking: 88% of healthcare leaders expect their AI agents to operate autonomously to some degree, but nearly three-quarters admit those same agents are going live without the IT department signing off first.