Physicians' 86% privacy condition makes clinical AI a processing question, not a model question

Eighty-six percent of physicians told the American Medical Association's 2026 Physician Survey on Augmented Intelligence that data privacy assurances would facilitate AI adoption in clinical practice ^[1]. In the same survey, 88% said validation that AI tools are safe and effective is equally important ^[2], which places the binding constraint on clinical AI programmes somewhere other than the leaderboard you picked your model from.

The gap between the two figures is two percentage points, with validation slightly ahead of privacy ^[3]. That ordering matters for anyone building a business case. These are not alternative objections to be traded off; they are two separate gates, and a programme that clears one still stops at the other. Privacy assurance is largely an infrastructure and contracting problem. Safety validation is a clinical evidence problem with a different owner, a different timeline and a different budget line.

On the privacy side, the argument being pushed at technology leaders is that the protection has to extend to data in use. John Pettit, chief technology officer of Promevo ^[5], writing in a Forbes Tech Council post ^[6], argues that confidential computing should be a priority for every healthcare technology leader building with AI ^[10], on the grounds that traditional security measures protect data only at rest or in transit while confidential computing safeguards information while it is being processed ^[4]. His case for urgency is that the next generation of AI often requires securely analyzing information from multiple sources ^[7], while much patient information remains fragmented across providers and health organizations, with boundaries that protect privacy but also limit the comprehensiveness of the insight ^[8]. The worked example is a patient who falls ill after travelling: the treating physician has the medical history, population health data elsewhere could reveal an emerging drug-resistant pathogen where the patient travelled, and without a secure way to analyse both together clinicians may fall back on broader treatments while waiting for tests ^[9]. Pettit's claim is that a trusted computing foundation removes a major barrier to healthcare AI ^[12].

Two caveats are worth keeping in front of the board. The prescription arrives from a supplier-side executive in a sponsored council slot ^[5][6], and the vendor conclusion follows the survey data more tightly than the survey data compels it. And the survey figures are quoted without sample size, fielding dates or question wording ^[11], so 86% is a stated condition for adoption, not evidence that privacy is the largest single blocker or that any particular technology satisfies it. Stated preference in a physician survey is not a signed order from a health system.

What changes in practice is the artefact a CIO or chief medical information officer owes their governance committee. It is no longer a model comparison; it is a map of where protected health information is physically processed during inference, and which supplier will commit in writing that the data stays protected while in use ^[4].

Watch whether the AMA publishes the methodology behind the two figures, because the 88% validation requirement ^[2] will need clinical evidence that no infrastructure purchase supplies. Watch whether vendors answer the in-use processing question in contract language rather than in bylined posts. And watch which of the two gates actually receives funding in 2026 budgets, since the survey says both are conditions of adoption ^[1][2].