Research suggests that doctors trust AI over evidence

According to the American Medical Association, more than 80% of doctors use AI. They leverage it when determining everything from standards of care (39%) to diagnosis (17%). The technology is now ubiquitous in the field – and likely has many benefits for both physicians and patients.  

Still, over-reliance on AI can lead to both overdiagnosis and underdiagnosis. A 2025 literature review included a number of striking cases. For example, underrepresented subgroups were ignored in melanoma diagnoses. Benign lung nodules were flagged as malignant. And perhaps most alarmingly, the authors cited a study that found 41% slower error identification in workflows that used AI versus those that relied solely on human judgment.  

A July 2026 study published in PLOS Digital Health used a series of digital tests to investigate just how reliant doctors were on AI determinations. The results were damning. Even when presented with results that contradicted AI, doctors were still inclined to trust machine reasoning over their own analysis.  

The study established two scenarios that required medical professionals to assess the efficacy of a treatment for a fictitious disease on groups of patients incorrectly classified by AI. In both, the AI classified patients as being highly or lowly sensitive to the treatment. 

In the first experiment, the doctors observed the AI recommendations – recommending more doses of medication to those patients classified as highly sensitive. The treatment, however, showed medium efficacy in both the highly and lowly sensitive groups.   

When asked to judge the efficacy of the treatment in practice, the doctors maintained that their adherence to the AI recommendations was mostly correct.  

“They judged the treatment to be highly effective for patients classified as highly sensitive by the AI. They judged the treatment to be moderately effective for patients classified as lowly sensitive,” said lead author Aranzazu Viñas, a psychologist with the Department of Economics and Management, University of the Basque Country, Spain. 

They did so even when presented with evidence to the contrary. Even more alarmingly, they made the same mistake when the treatment was not effective at all. 

“In Experiment 2, the treatment was completely ineffective,” Viñas adds. “That is, the healing rate in Experiment 2 was exactly the same when the treatment was administered and when it was not.” 

The doctors did not pick up on this even when presented with evidence that their patients had not healed. 

These results closely mirrored those of an earlier study conducted by the authors using lay subjects. That is, doctors appeared no more proficient at identifying AI classification errors than non-professionals. Viñas attributes this in part to causal illusion.  

“The causal illusion is a cognitive bias consisting of believing that two events are causally connected when in fact they are not,” she says. “The causal illusion also seems to affect professionals in their field, such as doctors faced with a patient classification system.” 

That is: doctors are inclined to trust preconceived notions rather than apply critical thinking to the results obtained. In essence, it is professional laziness. The study acknowledges that this effect may be less pronounced in real life – particularly in specialist fields that demand constant adjustment of treatment. However, they suspect that the tendency to trust AI may be of real concern in general practice, which is heavily reliant on broad classification systems and generalized protocols.  

Patient classification is a necessary part of medicine, both in terms of treatment and allocation of sometimes scarce resources. The Manchester Triage System, for example, categorizes emergency patients according to urgency.  

Even outside of formal systems, classification occurs unconsciously and unintentionally – sometimes resulting in deleterious effects due to discrimination, as Viñas notes. If doctors become more reliant on classification by AI – and ignore factors that contradict those classifications in their own examinations – patients could be at risk.  

“This tendency might be overcome by developing strategies and protocols that increase human critical thinking and detection of AI errors, in order to maximize the benefits of the human-AI collaboration while minimizing potential errors,” she suggests. “Specifically, training in critical thinking has proved effective in reducing the causal illusion.”  

Viñas echoes wider concerns about the use of AI in medical settings. While it has proven useful in everything from image analysis to drug discovery, guardrails are provisional at best. 

“What I find worrying is that we are using AI without any prior training, without fully understanding its pros and cons, how it affects human decision-making, under what conditions it can improve our decisions and under what conditions its use is inadvisable,” she cautions.