That scenario anchors a multidisciplinary Policy and Practice Review published in Frontiers in Oncology, in which clinicians from breast surgery, psychiatry, anaesthesia and perioperative medicine argue that general-purpose conversational AI has created a new layer of cancer care without creating a new accountable care system. The authors, led by Zacharoula Sidiropoulou of the West Lisbon Local Health Unit and NOVA Medical School, conceptualise patient-initiated AI use as a “parallel pathway”: an unrecorded sequence of questions, answers and actions that runs alongside formal care, is usually invisible to the treating team, and re-enters the clinic only through its downstream consequences — a fixed surgical preference, a delayed call about a fever, a quietly missed dose of endocrine therapy.

The scale of the phenomenon is still poorly characterised. A 2025 survey of 2,002 adults in the United States found that 32.6 percent had used a large language model for a health query, though oncology-specific prevalence was not measured. A 2026 cross-sectional study of 400 people with cancer and caregivers reported that 24 percent had used generative AI for cancer-related information; strikingly, 69 percent of those users had not disclosed this to a physician, while 73 percent of all participants said they would be willing to use AI if it were physician-verified. Most existing evaluations test whether an AI answer is accurate when responding to standardised questions. They cannot show how a frightened patient frames a prompt at midnight, how several answers accumulate across a conversation, or whether a question simply disappears from the next clinical encounter.

The review’s central conceptual move is to distinguish model error from pathway error. A model error is a false, fabricated or misleading output. A pathway error occurs when an exchange — even one containing no factual mistake — produces an unsafe clinical consequence. A factually correct statement can become dangerous if it is transformed into an unsupported individual prediction, a reason not to call the hospital, or a clinical instruction stripped of the conditions that make it safe. Conversely, an inaccurate answer may cause no harm if the patient rejects it or brings it promptly to the team. The relevant unit of safety analysis, the authors argue, is therefore not the isolated answer but the entire sequence linking an answer to behaviour and back to care.

Several mechanisms make this conversion plausible. Conversational systems display what the authors call interactional authority: they mirror the patient’s language, remember prior turns and answer without visible fatigue, producing what feels like individual understanding even when decisive clinical variables are absent. Experimental work shows that a system’s drive to be helpful can compete with its ability to correct false premises embedded in medical questions. There is also recursive preference construction: a patient begins with a question, receives a provisional answer, reveals selective additional details and asks the system to refine the emerging conclusion, with each turn becoming part of the next input. Notably, the technology can also behave protectively — in an evaluation of 908 authentic medical queries, 97 percent of responses advised consulting a professional — but referral language is not the same as successful escalation, and it can be diluted by a long, reassuring answer.

The review maps four distinct failure modes across the breast cancer pathway. In surgical decision-making, the risk is premature preference closure: a chatbot may state the population-level principle that long-term randomised trials found no overall survival advantage for mastectomy over breast-conserving surgery plus radiotherapy, yet apply it to a patient whose disease distribution, radiotherapy options or genetic result it does not know. Contralateral prophylactic mastectomy illustrates “precision without a valid denominator” — a precise percentage of future contralateral risk that appears individual because it is expressed exactly, even though the variables needed to individualise it, such as age, inherited susceptibility and tumour biology, were never supplied. Professional consensus discourages routine contralateral prophylactic mastectomy for average-risk unilateral cancer.

In treatment toxicity, the decisive question is not whether the AI answer was correct but what it did to the patient’s threshold for making the call. The authors present illustrative composite scenarios, explicitly labelled as hypothesis-generating teaching constructs rather than estimates of event frequency. In one, a woman ten days after her first chemotherapy cycle records a temperature of 38.1 °C and describes it to a conversational assistant that knows nothing of her regimen or her expected neutrophil nadir; advised to hydrate, rest and seek review only if she deteriorates, she presents eighteen hours later hypotensive, with grade 4 neutropenia and Gram-negative bacteraemia. Suspected neutropenic sepsis is an acute medical emergency requiring immediate assessment and empiric antibiotics — context a general-purpose system cannot be assumed to possess. In another scenario, a patient receiving a checkpoint inhibitor is told by an AI to stop treatment until her diarrhoea settles, moving a systemic therapy decision outside the accountable clinical pathway.

Psycho-oncology and perioperative care reveal subtler risks. A randomised trial of 150 patients with newly diagnosed solid tumours found that ChatGPT-assisted counselling was associated with lower anxiety and depression before the second chemotherapy cycle, and patients used it to discuss concerns they had never raised with clinicians — evidence of genuine supportive potential. But the same continuous availability and apparent empathy can foster repetitive reassurance-seeking, in which distress appears resolved in clinic while the underlying cycle of anxiety, reassurance and renewed anxiety remains active outside care. In one scenario, a woman in remission consults AI several times a day asking whether a pain means her cancer has returned, reformulating the question until she receives comfort. In the perioperative domain, the problem is loss of conditionality: generic fasting rules, medication advice and exercise programmes may be defensible as population guidance yet unsafe without knowledge of anaemia, thrombocytopenia, drains, wound status or a specific anaesthetic plan. A postoperative patient who describes a reddening wound and mild fever to a chatbot, and is told this can represent normal healing, may wait while an implant infection progresses — the assessment that mattered required someone to examine the wound, and the decision to observe came from a system that could not observe.

The authors propose a five-step clinical response — ASK, RECONSTRUCT, CLASSIFY, CHECK and RETURN — built around a routine, non-judgemental question: “Have you asked an AI about this?” The clinician asks what conclusion the patient drew and what action followed, classifies the influential output as general education, individual prediction or instruction, identifies which verified clinical variables were unavailable to the system, and returns the decision to accountable care through reopened deliberation, teach-back and proportionate documentation. A complete transcript is rarely necessary; recording the material belief, any action taken and the corrected plan limits unnecessary capture of sensitive data. On governance, the review argues that regulation such as the EU AI Act, whose main high-risk obligations apply from December 2027 and August 2028 following the Digital Omnibus, can govern products and deployed systems but cannot by itself reveal patient-initiated consumer use, because a consumer chatbot does not become medical-device software merely because a patient asks it whether mastectomy is safer.

The framework comes with candid limitations. It is a targeted narrative synthesis rather than a systematic review, with no formal risk-of-bias appraisal; breast-cancer-specific evidence on patient-initiated AI use and its behavioural consequences remains sparse, so some propositions rest on mixed oncology populations, model-output evaluations and adjacent clinical domains. The proposed safety properties — traceability, reversibility and returnability — are research constructs, not validated measures, and the five-step response has not been prospectively tested. The authors are explicit that their mechanisms are hypotheses for evaluation, not established dominant patterns, and that the technology itself is changing too rapidly for findings from individual model evaluations to generalise across platforms and versions. Their conclusion is neither prohibition nor endorsement. Conversational AI may prepare, explain, rehearse and reinforce; it should not close deliberation, close escalation, replace acute psychological assessment or originate instructions that depend on findings it cannot observe. The most immediate place to start, they suggest, is a single question that makes the invisible pathway visible.

Subject of Research: Patient-initiated conversational AI use between appointments in breast cancer care and its proposed clinical and governance responses

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

Article References: Sidiropoulou, Z., da Ponte, G., Kumar, S., & Fonseca, V. (2026). Patient-initiated conversational AI between appointments in breast cancer care: a multidisciplinary clinical and policy perspective. Frontiers in Oncology, 16, Article 1981490. https://doi.org/10.3389/fonc.2026.1981490

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1981490

Keywords: breast cancer, conversational AI, large language models, patient safety, shared decision-making, treatment toxicity, psycho-oncology, perioperative care, health governance, medical ethics, cancer care pathways, AI regulation