The two figures in that January survey move in opposite directions. That gap is where a deployment plan runs into trouble. The fall in openness works out to ten points in two years, a relative decline of about 19% [1][2]. Meanwhile the share of adults who report having made an important health decision with AI alone sits nine points above the share who report openness to AI in healthcare at all [3]. Use and consent are separate variables, and reported use does not predict reported openness.
The overshoot is mechanical. The five cognitive tasks in the JAMA argument are measured as tasks [2], and the studies underneath the debate are largely simulations of single tasks rather than real encounters [5]. Roy Lirov, a physician who also builds his own AI infrastructure, put the measurement problem plainly: "I think we're fooling ourselves by reporting on what's easily measured instead of figuring out how to measure what matters" [8]. A benchmark measures whether a task can be completed. The number that actually sizes a clinic is the share of encounters a patient will let you convert.
A skeptic reads the same data the other way: stated openness is soft, it follows exposure, and the figure on solo AI health decisions shows behaviour already running ahead of attitude. That reading is available, but it has to survive the fact that the two lines moved apart rather than together over two years [3][4]. The Forbes account also gives the headline percentages without sample size or question wording [15], so what the record supports is direction rather than magnitude, and on timing the column is blunt that nobody knows when AI will be ready to remove a doctor from the exam room [14].
The board-deck version is coherent. Workforce shortages and record physician burnout on one side, five tasks where the machine is said to match or beat the clinician on the other, therefore a staffing model with fewer clinicians in it [12][2]. What the deck omits is the distribution question the column raises, which is whether this care reaches patients broadly or mainly the ones already wearing an Oura ring or an Apple Watch [13]. A plan that works only for the instrumented and the willing is a smaller plan than the shortage figures imply.
Here is the trade-off. Buying documentation relief and data gathering, which the column says already works and which Lirov identifies as AI's best current use, produces cost reduction this year and a longer trust record for later [10][7]. Sizing a plan to a 2030 substitution thesis [1] spends trust that has not been accumulated, in front of patients whose reported openness is falling. The thirteen clinicians who told the author replacement is not possible are a convenience sample rather than evidence about 2030 [9], though they are also the people who would have to hand over the encounter. Capacity built to a benchmark still costs money if patients refuse it.
