Should Cardiologists be Concerned About Large Language Model–Based Chatbots From a Patient‐Complaint Perspective?

If you know how to use artificial intelligence (AI), including large language models (LLMs), it can save you, as a clinician, enormous time. ¹ Reviewing a complex referral? Upload a deidentified copy for a concise summary and key problems list. Facing a backlog of results? Let AI draft templated follow‐up notes you can verify and sign. Need a quick refresher on new guidelines? Feed a recent consensus statement or trial PDF to a chatbot and get a short synopsis plus practice points. Educating patients? Ask AI to generate bilingual handouts tailored to literacy level.

As these examples suggest, the implications extend beyond convenience. As AI tools become routine in clinical practice, they may narrow one of health care's most persistent asymmetries: the information advantage held by clinicians over patients. When used appropriately, patient‐facing AI tools may help patients understand medical terminology, interpret test results and reports, prepare for consultations, and participate more effectively in shared decision‐making. With decision aids that translate evidence, clarify options, and estimate individualized risks and costs, patients may be less susceptible to low‐value care, confusing billing, or steerage toward unnecessary services, potentially improving outcomes and system efficiency ² , ³ . Greater transparency is around the corner.

Information asymmetry has existed for as long as medicine itself. Historically, patients lacked the means to verify diagnoses or assess the quality and appropriateness of treatments and procedures. When the recipient of a service cannot readily judge its quality, they may assume the best, trusting the authority symbolized by the white coat. Yet in recent years, the medical world has become increasingly associated with issues such as financialization, rising insurance premiums, delays in care, and medication shortages. In health care, such dynamics can erode trust, lower overall quality, and ultimately leave more patients' needs unmet; and, let's be frank, can even fuel suspicion.

The internet has made it easier for patients to get informed. Patient forums and communities translate medical jargon into plain language and offer first‐hand accounts of surgeries, cancer care, and hospital journeys. Plain‐English glossaries and explainer sites help demystify diagnoses and treatments. Ratings platforms, such as Google Reviews and dedicated health care review sites, now let patients evaluate clinicians and hospitals, nudging providers to improve communication, empathy, and overall quality. Little by little, sunlight is squeezing out the worst of information asymmetry in medicine.

Information perfection is on the rise. Today, AI is expected to deliver us from opacity, confusion, and inertia. And the medical world is no exception. As one observer put it, AI “will help people who didn't have the privilege of great advice to get … pretty great advice.”

One contemporary stereotype is that some patients, particularly those in Generation Z, may be reluctant to accept a prescription or recommendation without first reviewing their medical report with tools such as ChatGPT, Claude, or Gemini. This can sometimes sharpen questions and improve understanding, but it can also go badly wrong. Ask an AI, “Can beta blockers and flecainide be taken together?” and you may suddenly find a tranche of patients with atrial fibrillation deciding (wrongly) not to take their medication.

So, should clinicians fear AI from a complaint point of view? The answer is: only partly, and it depends on 4 things (Figure). These 4 domains are not only conceptual; they can be translated into near‐term actions at the point of care, within local health‐system governance, and through postdeployment monitoring.

Figure 1. In the current era of ubiquitous AI‐driven health information, patients increasingly arrive at cardiovascular consultations having already sought guidance from large language models.

This conceptual figure depicts the emerging profile of the AI‐assisted patient: partly informed yet potentially misinformed, having compared AI‐generated recommendations directly against anticipated clinical advice. The resulting informational dissonance generates conflicting answers that drive doubt and anxiety, heightens expectations for idealized guideline‐concordant care, and increases willingness to question or challenge clinical judgment. The cumulative downstream effects on the patient–clinician relationship include greater disagreement, dissatisfaction, and an elevated risk of potential complaints. AI indicates artificial intelligence.

FIRST: PATIENT AI LITERACY

AI can improve care only if patients know how to use it well and understand its limitations. Simply parroting a chatbot's output (“it said this drug is dangerous”) is far less useful than using AI to understand options, risks, and guideline logic, so patients can ask better questions. Poor use of AI will generate noise, anxiety, and potentially nonadherence, all of which land back on the clinician's desk. In practice, this means health systems and professional societies may need to offer simple patient‐facing guidance: what health chatbots are useful for, what they are not reliable for, what information should never be entered, and when direct clinical review is necessary. A “health chatbot users' guide” may prove more valuable than abstract reassurance alone. ⁴

SECOND: RESPONSIBLE CLINICAL INTEGRATION

Clinicians and health systems will have to show how AI was used: what was recommended, what was ignored, and why. That means transparent outputs, checks against clinical context, and workflows that keep human judgment in charge. If something goes wrong, a defensible trail (“the AI suggested X, clinician chose Y because …”) will matter for complaints and for insurers. Operationally, this could be as simple as routinely asking whether a chatbot materially shaped the patient's interpretation, expectations, or treatment decisions; explicitly reconciling major discrepancies between chatbot output and the clinical plan; and document the reason for that difference when it materially affects care. ⁵

THIRD: GUIDELINE TRANSPARENCY AND TOUGHER CONVERSATIONS

Patients will increasingly quote guidelines back to clinicians: “The 2025 guideline gives a Class I recommendation for X, but you prescribed Y. Why?” That will make some encounters more demanding. It also means clinicians must stay current at a pace that, for many, used to be more aspirational than operational. In cardiovascular medicine especially, this challenge is sharpened by the rise of AI‐enabled, increasingly dynamic, and potentially “living” recommendations, ⁶ in which evidence, thresholds, and recommendation strength may evolve faster than traditional guideline cycles. As patients arrive with AI‐generated summaries of guideline‐based care, the room for unexplained variation narrows. The practical response is not defensive medicine, but clearer explanation: clinicians will need concise ways to explain not only why a recommendation fits the patient in front of them but also when comorbidity, frailty, drug interactions, patient preference, incomplete context, or rapidly evolving evidence justify a different course from an apparently authoritative AI‐generated answer. In some cases, the harder conversation will no longer be whether the clinician knows the guideline, but whether they can explain when formal guidance is still catching up with the evidence base. Whether this will translate into more complaints, litigation, or malpractice‐premium adjustments is not yet established by robust trend data. However, it is a plausible scenario that warrants attention, particularly in health systems and private‐practice environments where complaint signals and insurer responses may emerge quickly.

FOURTH: REDEFINING MEDICAL RESPONSIBILITY

Responsibility will have to be explicitly allocated. Can a clinician “share” liability with an AI tool? If so, under which standard? Complaints bodies and courts will start asking: Who made the decision: the doctor, the hospital's AI, or the patient's chatbot? The fuzzier that is, the more it falls back on the human clinician. For that reason, institutions will need more than clinician vigilance alone: they will need visible governance around which tools are acceptable, how outputs are validated locally, what disclosures are made to patients, and how complaints, near misses, and workflow failures are monitored after deployment. In other words, oversight must extend beyond the bedside to a clinical loop, a governance loop, and a learning loop. ⁷

Because cardiovascular care is guideline‐rich, high‐stakes, and increasingly digital, cardiology may offer an early and particularly visible setting in which to observe these frictions. A related shift has already occurred on the professional side. Tools like OpenEvidence, marketed as the fastest‐growing physician app, integrated with leading peer‐reviewed sources (JAMA, New England Journal of Medicine, etc), and raised sizable rounds, including a $210 million raise at a $3.5 billion valuation in July 2025, reportedly reaching about $6 billion by November 2025. With AI surfacing differential diagnoses and treatment options in seconds, the clinician's center of gravity will move from “Can I spot the diagnosis?” to “Can I document, explain, consent, and follow up lawfully?”

AI may shorten parts of diagnostic workups or, at minimum, create the expectation that this should happen more quickly. But adverse events, interactions, social factors, and imperfect adherence still happen. So when a patient feels, “I followed the AI, I followed the doctor, and I was still harmed,” the blame is more likely to land on the human part of the chain. The part that was supposed to contextualize, warn, and document.

A critical reader may—rightly—ask for stronger evidence and more granular references on the medico‐legal implications, the physician–patient relationship, and health‐literacy pathways that underpin our intentionally provocative claim. That request is fair. Multiple data sets ⁸ , ⁹ , ¹⁰ suggest that LLMs can lower informational barriers by translating, simplifying, and explaining medical information. These potential benefits deserve explicit acknowledgment, particularly for patients trying to understand unfamiliar terminology, laboratory results, imaging reports, or treatment options. However, these examples assume an “optimal‐use” scenario: well‐crafted prompts, cautious interpretation, and complete clinical context. At the same time, the evidence base remains less mature than the pace of adoption might suggest. A recent Nature Medicine review identified 4609 clinical LLM studies published between January 2022 and September 2025 (roughly 3.2 papers per day), yet only 1048 used real‐world patient data and only 19 were prospective randomized trials. The same review also showed that most studies relied on simulated scenarios or exam‐style tasks rather than real‐world clinical data and that at least 25% had sample sizes <30. ¹¹ In parallel, across 31 LLMs, the potential for severe harm from LLM‐generated medical recommendations has been reported in up to 22.2% of cases. ¹² Together, these data suggest that the field is expanding faster in volume than in the strength and maturity of its evidence base.

From a patient‐complaint perspective, the central issue is real‐world use: rough adoption, wide variation in health literacy, missing or distorted context, improperly calibrated trust in outputs, and the emotional charge that accompanies interpretation of health information (anxiety, reassurance‐seeking, fear). In other words, the relevant question is not whether AI can perform well under ideal benchmark conditions but how it will be used by real patients in real clinical pathways. The challenge is that practice is evolving faster than the scholarly record. In early December 2025, the New York Times ¹³ reported that patients were already feeding blood tests, clinicians' notes, and operative reports into general‐purpose chatbots. On January 7, 2026, OpenAI announced ChatGPT Health, a dedicated health experience inside ChatGPT designed to support—rather than replace—clinical care, including the option to connect personal health data so responses can be more personalized and grounded in context. This sequence crystallizes a central point: large‐scale, patient‐driven use of LLMs for medical‐record interpretation and health conversations is no longer hypothetical. It is already material enough to shape product strategy and risk exposure, raising concrete questions about health‐data sovereignty, cybersecurity and privacy safeguards, and downstream clinical harm from inaccurate or hallucinated outputs. These concerns extend beyond output quality to data governance. When patients upload medical records, laboratory results, imaging reports, or other sensitive information into third‐party LLM platforms, questions arise regarding data ownership, secondary use, storage location, cross‐border transfer, and how such data may be retained, shared, or repurposed. Privacy and cybersecurity are therefore not peripheral considerations, but central elements of the broader framework of accountability, public trust, and societal confidence in the use of patient‐facing AI in clinical practice.

Responsible deployment should treat patients as stakeholders alongside clinicians, researchers, policymakers, and institutions, by investing in patient AI literacy, making uncertainty and provenance transparent, and protecting time and space for empathic explanation and shared decision‐making. Yet the current pace of adoption leaves little room for this ideal trajectory. In the near term, medico‐legal risk could increase as uncertainty, accountability gaps, and rapid adoption converge. However, the magnitude and timing of this shift remain uncertain and should be framed as a plausible scenario rather than as an established trend.

Sources of Funding

This research was supported by GERCA (Groupe pour l'Enseignement et la Recherche Cardiovasculaire en Alsace, Strasbourg, France), a charitable cardiovascular research consortium.

Disclosures

Benjamin Marchandot, Antonin Trimaille, and Olivier Morel declare that they have no competing interests, whether financial or non‐financial, that could be perceived as relevant to the manuscript. Olivier Morel reports grants in support of investigator and investigator‐initiated studies from AstraZeneca, Medtronic and Boehringer Ingelheim, all outside of the submitted work. Olivier Morel has been awarded grants by “Fondation Cœur et Recherche” and “Endofrance,” reputable charities in France committed to advancing research initiatives in cardiovascular disease in endometriosis. Antonin Trimaille acknowledges Fédération Française de Cardiologie, Groupe Insuffisance Cardiaque et Cardiomyopathies of the French Society of Cardiology, and Villa‐M for their support during his year in Canada as a clinical research fellow. Benjamin Marchandot has no disclosures to report.