Doctors tripled their AI use — and tripled their list of limits

The doctors and nurses who have integrated AI into daily practice are the ones most resistant to expanding its role. A survey of 355 US physicians and nurses commissioned by Wolters Kluwer and conducted in March 2026 found that clinicians’ daily AI use tripled over the previous year. In the same sample: 74 percent fear the technology will erode the clinical skills it is replacing, 74 percent distrust its outputs because of hallucinations, and 72 percent are concerned that advertiser-driven business models will distort medical recommendations. The people using AI most are drawing the boundary most firmly.

What that boundary looks like in practice: clinical imaging and diagnostics sit on the acceptable side. Documentation, treatment recommendations, patient communication, and anything touching clinical judgment sit on the other. Clinicians have spent years watching AI improve at reading scans and flagging anomalies — improvements that are measurable, published, and peer-reviewed. The same evidentiary framework does not exist for the broader applications that developers are now pushing into hospitals. Doctors are not rejecting AI. They are applying the same scrutiny to it that they apply to any clinical intervention.

The governance gap is measurable and is not closing quickly. The share of clinicians who say they are aware of AI governance policies at their institution rose from 21 percent in 2025 to 27 percent last year. That is a six-point gain on a figure that should be near 100. When institutions deploy AI systems for clinical use without corresponding governance frameworks, clinicians fill the gap themselves — which is why individual resistance is highest in the areas where institutional oversight is weakest.

The hallucination concern is not theoretical. Large language models used in clinical contexts have produced plausible-sounding but factually incorrect information about drug interactions, dosages, and diagnoses. The concern about advertiser-driven bias reflects something more structural: most clinical AI tools are built by companies whose revenue depends on engagement and upsell, not on patient outcomes. Clinicians working inside those systems recognise the incentive misalignment even when vendors do not advertise it.

What the survey describes is not technophobia. It is a profession applying evidence standards consistently. Radiology AI passed because it generated a body of peer-reviewed evidence and submitted to clinical trials. The tools now being deployed for documentation, triage, and patient interaction have not gone through that process at comparable scale. The Financial Times, reporting on the survey today, frames this as pushback. It is more accurately a request for the same evidentiary bar every other medical intervention has to clear.

The 2026 Stanford HAI AI Index tracks a similar pattern in its medicine section: adoption is running ahead of validation, and the places where clinical AI is most embedded are rarely the same places where the outcomes literature is established. What the Wolters Kluwer data adds is the practitioner perspective — that the people closest to the technology are the ones articulating most clearly where the evidence runs out.

Tags: artificial intelligence, clinicians, deskilling, diagnostics, healthcare, medical AI