Patient-initiated conversational AI between appointments in breast cancer care: a multidisciplinary clinical and policy perspective

General-purpose conversational artificial intelligence (AI) is increasingly used for health-related questions, including questions relevant to cancer care, although breast-cancer-specific prevalence and behavioural effects remain poorly characterised. Most evaluations ask whether an AI answer is accurate. That is necessary but incomplete. Clinically consequential effects may arise when a fluent, apparently personalised exchange changes a treatment preference, delays contact with the clinical team, reinforces reassurance-seeking, validates non-adherence, or converts population guidance into an individual-sounding instruction. We conceptualise this as a parallel pathway: an AI-mediated sequence of questions, answers and actions that is usually invisible to the treating team, absent from the medical record and unattached to professional responsibility, yet capable of re-entering formal care through its downstream consequences. We distinguish model error from pathway error and propose traceability, reversibility, and returnability to accountable care as safety properties beyond answer accuracy. Across four clinical domains, we propose distinct pathway failure modes: premature preference closure in surgical decision-making; escalation closure during treatment toxicity; repetitive reassurance-seeking and over-reliance in psycho-oncology; and loss of conditionality in perioperative recovery and lifestyle advice. These are hypothesis-generating mechanisms rather than empirically established dominant patterns. We propose a five-step clinical response - ASK, RECONSTRUCT, CLASSIFY, CHECK and RETURN - based on non-judgemental inquiry, proportionate documentation and explicit return-to-care boundaries. Regulation can govern products and deployed systems, but it cannot by itself reveal patient-initiated consumer AI use. Clinical services therefore need governance for the pathway as well as for the technology.