2,441 studies on generative AI in cancer reviewed: none asked patients about safety or autonomy

By Zendoric · September 5, 2026.

A team led by Flinders University (Australia) has carried out an unusual exercise: instead of measuring how accurate generative AI is in oncology, it has measured how much the people on the receiving end have been asked. The result, published in the European Journal of Cancer (DOI: 10.1016/j.ejca.2026.116990), is a scoping review that screened 2,441 scientific records and found that only 32 studies directly captured the perspective of cancer patients, survivors or caregivers on the use of generative AI in their care.

Within those 32 studies, according to the authors themselves, the focus was narrow: most assessed only whether generative AI improved the accessibility and clarity of information about cancer, or whether it simplified complex clinical reports so that patients could understand them. On that front, the reception was generally favorable. But accountability (who answers if the AI fails) was addressed by just three studies. And none —zero— assessed what patients think about safety, transparency, equity or the preservation of human autonomy in the face of these tools. The study also notes that patient trust depended on whether the AI's answers were perceived as personalized, emotionally appropriate and accurate, not merely factually correct.

"Designing AI for patients is not the same as designing it with them," says Bradley Menz, a Flinders researcher and the paper's first author. If these tools are going to influence cancer treatment, he adds, patients, survivors and caregivers need a real role in how they are designed, governed and used. His colleague Ashley Hopkins points in the same direction: gathering those perspectives could improve everything from interface design to consent processes, integration into the clinical workflow and oversight and control mechanisms.

The finding matters because deployment is running ahead of the question. As sector context, hospitals and healthcare software vendors are already rolling out patient chatbots, automatic record summaries and assistants built on generative models, largely because the technology is available and promises to ease the administrative burden on clinical teams. What this review exposes is that the rollout is advancing without a systematic channel for finding out whether the patient —the most vulnerable person in the equation, often under extreme emotional stress— trusts the system, understands who answers for its errors, or feels they retain control over their own care.

It is a pattern we have already seen in other corners of the AI governance debate: the risk lies not only in the model's capability, but in who gets asked and about what. Here the blind spot is not technical but procedural: nobody has systematically asked about safety or autonomy because the design of these tools still starts from the provider's logic —clinicians, developers, hospitals— and not from that of the person who will use them at the most fragile moment of their life.

Our take is that this gap is fixable, and that is the hopeful part: it is a problem of research methodology and governance, not of the technology itself. The same generative AI that today does little more than translate clinical jargon into plain language has, in the medium term, real potential to speed up screening, personalize treatment information and free up oncologists' time for what no machine can replace: the difficult conversation, the support, the clinical judgment in an atypical case. That is precisely the ground on which AI can bring us closer to treating cancer as a manageable chronic disease rather than a death sentence, part of the health and longevity horizon we defend as our underlying thesis. But that long-term promise only holds if trust is built starting now, including those who are currently left out of the design. A review that documents silences, like this one, is the necessary first step toward filling them before adoption becomes widespread without the right question having been asked.