What Cancer Patients Actually Think About AI: A Comfort Gradient, Not a Verdict

Across eight clinical tasks, oncology patients' comfort with artificial intelligence spans about nineteen percentage points — from 80.2 percent at one end to 61.5 percent at the other. The two ends are the point. The high mark is cancer screening, the kind of result you can repeat or overturn. The low mark is prognosis, the estimate that tells you how much of a life is left. Comfort does not collapse between them; it slides, task by task, in step with how much each decision actually costs.

That slope is the whole story of Patients' Attitudes and Beliefs Toward Artificial Intelligence Use in Cancer Care, a survey from the Memorial Sloan Kettering Cancer Center published in JMIR Cancer. Its value is partly in who it asked. The literature on AI in oncology is overwhelmingly the voice of the people who build, regulate or deploy the systems; the patients on the receiving end are rarely polled in any structured way. This study polls them.

The design, and its ceiling

It is a cross-sectional survey — a single snapshot, not a trial, not a cohort tracked over time. Between September and December 2024 the team approached 383 adult oncology patients in clinic; 330 completed the questionnaire, an unusually high 86.2 percent response rate. Patients rated their comfort with AI across eight clinical touchpoints on a five-point scale and their worry about specific harms on a three-point scale. There is no intervention and no clinical endpoint anywhere in it. What it captures is stated attitude at one moment — genuinely useful at that strength, and easy to over-read past it.

The slope, and two surprises

The gradient is orderly. Comfort sat highest for screening (80.2 percent), then the supportive tasks — exercise advice (78.2 percent), diet (74.8 percent), herbs and supplements (72.4 percent) — and fell through the decisions at the heart of the disease: diagnosis (70.4 percent), symptom management (67.5 percent), treatment planning (64.8 percent) and, at the bottom, prognosis (61.5 percent). The logic is human, not technical. A screening flag can be checked and walked back; a prognosis cannot. Where the consequence is irreversible, patients want a person's hand on it.

Two associations are worth pausing on because they upset the usual assumptions. The first: familiarity, not ignorance, predicts calm. Patients who used web-based AI tools at least weekly were far less worried than less-frequent users (30 percent versus 53.6 percent; P=.001). The second cuts the other way. A university degree predicted more optimism that AI would improve care (83.5 percent versus 72.6 percent; P=.04), but left worry untouched (49.6 percent versus 49.3 percent; P=.97). Concern here is not an information gap waiting to be lectured shut. It is pointing at something the brochure does not address.

What it points at shows up in the second finding. Just under half the sample — 49.7 percent — voiced at least some concern about AI in their cancer care. The leading worry was neither algorithmic error nor a data breach. It was losing human interaction with their clinicians (49.1 percent), narrowly ahead of medical errors (47.9 percent), then privacy (42.7 percent), difficulty understanding the technology (40 percent) and the widening of healthcare disparities (33.9 percent). Read the order, not the individual numbers: the thing patients are most protective of is the relationship.

What the survey cannot do

Hold the result at arm's length for a moment, as the authors do. This is a convenience sample from one elite American cancer centre; 77.4 percent of respondents held a college degree, well above the national average, which the authors flag as a likely tilt toward favourable views. The questionnaire was administered in person, an arrangement that nudges answers toward what feels acceptable to say aloud. And because the design is cross-sectional, it cannot show how any of these attitudes shift as a patient moves through diagnosis, treatment and survivorship. None of these are the figures of a German or European cancer population — and they were never offered as such. There is a subtler ceiling too: a survey records comfort with the idea of AI, not the reaction when a specific algorithm's recommendation lands in a real consultation. Stated ease with 'AI in screening' is not acceptance of one model's verdict on one mammogram. The slope shows where goodwill runs thinnest; it does not predict how any concrete tool will be received.

Why it travels anyway

The numbers are local; the shape of the answer is not. The lesson is not that patients are for or against AI — half are quietly both at once. It is that their willingness is conditional and scales with the stakes, and that their first instinct is to protect the part technologists forget to cost: the doctor staying in the room. For any health system putting AI into oncology, that makes the path concrete and unglamorous. Trust is highest where the work is supportive and checkable, and it has to be earned rather than assumed as the tool edges toward diagnosis, treatment and prognosis. Whether patients in Germany and Europe would draw the same map is genuinely open — and reason enough to ask them rather than import the answer.

Source: Santos Teles M, Bryl K, Chimonas S, et al. Patients' Attitudes and Beliefs Toward Artificial Intelligence Use in Cancer Care: Cross-Sectional Survey Study. JMIR Cancer 2026;12:e81346. A single-centre, cross-sectional convenience-sample survey of stated attitudes; it measures perception at one point in time, not behaviour, outcomes, or a nationally representative population.