The ethics of listening walls: patient autonomy and consent in the age of ambient clinical AI

Abstract

Ambient clinical AI scribes are rapidly adopted to reduce documentation burden and clinician burnout. While evidence shows efficiency gains, these always-listening systems raise significant ethical concerns around informed consent, autonomy, privacy, data governance, and trust. This paper examines patient perspectives, identifies regulatory and equity gaps, and argues for patient-centered consent, transparency, data minimization, and robust governance to ensure ethical deployment of ambient AI in clinical care settings.

Introduction

Ambient clinical artificial intelligence (AI) systems—often referred to as “listening walls”—are rapidly being introduced into hospitals and outpatient clinics worldwide. These tools continuously capture conversations between patients and clinicians, processing them in real time to produce clinical notes, summaries, and structured documentation¹. For many health systems struggling with administrative burdens, ambient AI promises a much-needed transformation: reduced paperwork, improved documentation accuracy, and more face-to-face time with patients^2,3. Yet beneath this optimistic framing lies a set of profound ethical concerns^4–6. Continuous listening technologies are being deployed faster than clinicians, patients, or health regulators can fully understand them⁷.

Most importantly, the introduction of ambient AI challenges long-held ethical principles that underpin clinical practice—patient autonomy, informed consent, confidentiality, and trust. Many patients may not fully understand when or how these systems are recording, whether or how their data are stored, or who might eventually access the audio⁷. As ambient AI becomes normalized in clinical spaces, the ethical question is no longer whether we can deploy these systems, but whether we can do so responsibly. This paper argues that without robust consent mechanisms, transparent communication, and governance frameworks, ambient AI risks eroding the very foundations of ethical and patient-centered care it claims to support.

The rise of ambient AI in clinical documentation

The burden of clinical documentation is well established; evidence suggests that healthcare professionals spend an average of 2 h beyond their scheduled workday completing documentation tasks⁸. This administrative workload contributes to burnout and reduces time available for direct patient care. Additionally, variability in documentation quality—such as omissions of critical details or inconsistent formatting—can compromise patient outcomes and continuity of care⁹. Reducing documentation burden has therefore become a priority, both to improve clinician well-being and to enhance overall patient care¹⁰. These longstanding challenges have driven the search for more advanced technological solutions.

One emerging response is the use of ambient AI scribes, particularly those powered by generative AI and large language models¹¹. These systems rely on always-on microphones placed in examination rooms, on clinician devices, or embedded in clinical environments. They capture audio, filter out background noise, identify speakers, extract clinically relevant information, and generate structured documentation¹¹. Adoption is increasing. In a large cohort study of 1565 ambulatory physicians at an academic health system, 44.6% adopted ambient scribes within 2 years of their availability¹². Similarly, one large health system reported use by more than 7000 clinicians across over 2.5 million patient encounters within a 14-month period¹³.

Early evidence suggests that ambient AI scribes can reduce documentation time by 20–30%. While these efficiency gains create opportunities to improve clinician well-being and free up time for patient care, the extent to which such benefits are realized may depend on how time savings are allocated within healthcare systems^2,3. Empirical studies provide further support for these gains. For example, in a quality improvement study involving 45 clinicians from 17 specialties—including nurse practitioners and physician assistants—ambient AI scribes reduced documentation time by a median of 2.6 min per appointment and cut after-hours EHR work by 29.3%¹⁴. Similarly, an observational study of 119 allied health professionals found a 33% reduction in documentation time¹⁵. Beyond these efficiency gains, emerging evidence also suggests positive linguistic effects of ambient AI-generated documentation. A recent matched pre-post analysis of over 6000 outpatient clinical notes found that ambient AI systems produce more syntactically elaborated and structurally coherent clinical narratives, particularly in the History of Present Illness section, with increased sentence complexity and improved discourse coherence¹⁶. These findings indicate that ambient AI has the potential not only to reduce documentation burden but also to enhance the clarity, organization, and readability of clinical documentation, with likely benefits for communication and downstream clinical interpretation.

However, while these benefits are largely clinician-centered, the ethical risks fall disproportionately on patients. Ambient AI systems do more than transcribe: they capture patient voices, emotions, family interactions, and incidental conversations occurring in the room. They also generate vast amounts of data that may have secondary uses—whether anticipated or not—raising significant concerns about privacy, surveillance, and data governance⁷.

Ethical concerns raised by ambient clinical AI

The integration of AI into healthcare settings—particularly through ambient listening technologies—presents significant ethical and legal challenges, most notably concerning patient autonomy, informed consent, and data governance¹⁷. These systems, designed to record and analyze patient–clinician interactions, require patients to understand precisely what information is being collected, how it is processed, and the potential risks and benefits associated with its use. Meaningful consent depends on clarity about the technology’s functions, privacy implications, and the voluntary nature of participation¹⁷. Yet, current deployments of ambient AI often fall short. Subtle or poorly explained recording practices, limited opportunities to opt out, and the absence of clear alternatives can create environments where patients may feel pressured to accept AI-enabled documentation⁷. For vulnerable groups—including those with lower health literacy or limited understanding of AI—such conditions risk transforming consent into a form of implicit coercion rather than a genuine exercise of autonomy. These concerns are further intensified by the legal status of recorded patient–clinician interactions, which in many jurisdictions may constitute protected health information subject to established privacy frameworks (e.g., the Health Insurance Portability and Accountability Act (HIPAA) in the United States, the General Data Protection Regulation (GDPR) in the European Union, and analogous national data protection laws). Within such frameworks, the involvement of third-party technology vendors raises questions about the legal basis for data access, processing, and sharing, as well as the adequacy of consent mechanisms in meeting regulatory requirements.

Beyond initial consent, ambient AI technologies introduce complex uncertainties regarding data ownership, downstream use, and governance. Recordings and transcripts generated during clinical encounters are not merely byproducts of documentation but represent a distinct category of clinical data whose ownership and control remain poorly defined in practice⁷. It is often unclear whether such data belong to patients, clinicians, healthcare institutions, vendors, or some combination thereof, with existing arrangements frequently defaulting to institutional or vendor control without explicit patient-centered justification⁷. This ambiguity carries important legal implications, particularly in contexts where third-party companies process or store clinical data, raising concerns about accountability, stewardship, and regulatory compliance. At the same time, the large repositories of recorded conversations produced by AI scribes are increasingly used for algorithm development, model refinement, and commercial innovation⁷. Patients may be unaware that information shared in clinical encounters could contribute to future AI systems or proprietary products, and the absence of explicit, granular consent for such secondary uses raises concerns about trust, fairness, and potential exploitation¹⁷.

Closely related to these issues are concerns surrounding privacy, confidentiality, and legal accountability. The clinician–patient relationship has traditionally relied on assurances that sensitive information will be handled discreetly and securely¹⁸. Ambient AI systems complicate these expectations by capturing entire conversations—including potentially irrelevant or highly sensitive disclosures—and storing or transmitting them through cloud-based infrastructures managed by commercial vendors⁷. In many legal contexts, such recordings or their AI-generated transcripts may be treated as part of the formal clinical record and could be admissible in legal proceedings, such as malpractice litigation or professional disciplinary cases¹⁹. While this may enhance the evidentiary value of clinical documentation, it also raises concerns about how continuous recording may influence clinician behavior, potentially encouraging more defensive or performance-oriented communication at the expense of open and candid dialog¹⁹. The possibility of unintended data capture, unauthorized access, or extended retention further blurs established clinical boundaries and may contribute to the perception of clinical environments as monitored spaces, particularly for patients seeking care for stigmatized conditions.

These concerns intersect with broader issues of transparency and trust. Patients often lack clear visibility into what is being recorded, how long data are retained, who can access them, and whether they may be shared with third-party entities for development or commercial purposes¹⁷. Although some vendors, such as Amazon Web Services HealthScribe, have introduced features to enhance auditability, comprehensive transparency standards remain limited⁷. The opacity surrounding data processing, system performance, and governance arrangements contributes to uncertainty that may undermine confidence in both clinicians and healthcare institutions. Concerns about undisclosed secondary uses, including algorithm training or product development, further reinforce perceptions of imbalance between those who generate data and those who derive value from it¹⁷.

Taken together, these ethical and legal concerns underscore broader questions about clinical practice, equity, and regulatory oversight. Many AI scribes currently operate outside formal regulatory frameworks because they are classified as administrative rather than clinical tools, creating gaps in accountability, safety, and legal clarity⁷. While such technologies may enhance efficiency, they also introduce risks of documentation errors, insufficiently informed consent, and unequal performance across diverse patient populations²⁰. These dynamics highlight the importance of examining how ambient clinical AI may reshape power relations, redistribute risks, and influence the ethical foundations of care delivery in increasingly data-driven healthcare environments.

Privacy in healthcare versus consumer technologies

Initial public resistance to “always listening” consumer technologies—such as smart speakers and home security systems—has often given way to rapid normalization, raising the question of whether privacy concerns surrounding ambient AI in healthcare will follow a similar trajectory. In consumer contexts, users may tolerate continuous listening because the perceived benefits—such as convenience, automation, and personalized services—appear to outweigh the risks, and because data use is frequently framed as a matter of individual choice rather than a relational or institutional obligation. However, regulatory gaps persist in consumer settings, where health-related data from apps and wearables often fall outside traditional healthcare privacy frameworks, exposing users to secondary uses and commercial exploitation²¹.

Healthcare contexts, by contrast, are ethically and legally distinct. Clinical data are not merely transactional outputs but are embedded within relationships of trust, vulnerability, and professional responsibility. As such, they are typically governed by stronger regulatory regimes (e.g., HIPAA-style or GDPR-style frameworks) that treat privacy as a foundational condition of care rather than a negotiable feature. Empirical evidence further reinforces this distinction. Studies examining patient attitudes toward AI scribes indicate that privacy concerns do not dissipate quickly; many patients remain cautious despite expressing conditional trust, and awareness of AI use often remains limited even in settings where such systems are already deployed^22,23. These findings suggest that privacy expectations in healthcare are more resistant to normalization, particularly when recordings are directly linked to diagnosis, treatment, and deeply personal health experiences²².

Against this backdrop, it is difficult to assume that privacy concerns in healthcare will diminish simply through increased exposure or familiarity. The potential harms associated with ambient clinical AI—including erosion of trust, reduced disclosure of sensitive information, and the unintended commercialization of clinical interactions—are both more severe and more structurally embedded than those typically encountered in consumer technology settings⁷. Consequently, the ethical justification for robust safeguards in healthcare must rest not on expectations of user adaptation, but on the intrinsic moral significance of privacy within clinical relationships. Even if some degree of normalization occurs over time, the need for transparent consent processes, data minimization, and clearly defined limits on secondary data use remains non-negotiable, reflecting the distinctive ethical and institutional weight of privacy in healthcare⁷.

Evidence on patient perceptions and public attitudes

Existing literature reveals a complex interplay of acceptance and apprehension toward AI scribes. While some patients appreciate the potential for improved communication and increased clinician focus, others voice concerns about privacy, data security, and the loss of control over sensitive information²².

Pelletier et al.²⁴ evaluated the impact of a digital scribe in a pediatric pilot and reported a modest but statistically significant increase in provider-specific caregiver likelihood-to-recommend scores following implementation. In this study, “caregivers” referred to parents or guardians of pediatric patients. However, other patient-reported measures, including perceived provider listening and trust, did not significantly change. These findings indicate limited and mixed evidence of perceived benefit from caregivers, with improvements observed in overall recommendation ratings but not in other dimensions of care experience.

Chandrasekaran et al.²² analyzed data from an online survey of 12,153 Canadian adults (52.4% female; 23.1% ≥65 years; 41.2% with chronic conditions) and reported a complex pattern of attitudes toward AI scribes. Although 39.3% expressed some or high comfort with AI scribes and 49.5% anticipated positive effects on patient–provider interactions, 57.4% indicated they would only trust AI-generated documentation with human oversight, and 61.8% reported reluctance to use AI scribes in the future. Awareness of AI scribe use was also low (28.3%). Rather than reflecting a paradox, these findings may indicate conditional acceptance shaped by concerns about oversight, limited familiarity, and broader dynamics of deference within clinical relationships, where patients may hesitate to express full reservations that could affect their relationship with providers.

Evidence from experimental settings reinforces this complexity. A U.S.-based 2×3 factorial study—with two independent variables (health condition: acute vs. chronic; encounter type: AI substitution, AI augmentation, or traditional care), yielding six experimental conditions—involving 634 participants showed that attitudes toward AI in healthcare shifted significantly depending on both factors. Perceived privacy risks, trust issues, communication barriers, concerns about regulatory transparency, liability fears, expected benefits, and overall intention to use varied across the six scenarios, with misalignment between patient values and AI capabilities driving much of the resistance¹⁸.

Similar patterns appear elsewhere. In Australia, Evans et al.¹⁵ found that although most patients were comfortable consenting to the use of AI scribes during their own appointments, many believed that other patients would require additional information—particularly regarding data storage and security—before making an informed decision.

Overall, the evidence points to a clear conclusion: consumer acceptance of AI scribes depends heavily on transparency, control, and trust—elements that remain inconsistent across current deployments. These concerns underscore the need to strengthen transparency practices, protect patient autonomy, and establish robust trust-building mechanisms as AI-driven documentation tools continue to expand in healthcare.

Toward ethical deployment of ambient clinical AI

An ethically acceptable implementation process for ambient AI scribes requires more than technical installation; it demands a deliberate, multi-stage governance approach that explicitly protects patient autonomy, clinician professionalism, and justice in care delivery. This framework is informed by earlier healthcare information technology and mobile health (mHealth) tools, which showed that privacy protection, informed consent, data stewardship, fairness, and transparency must be addressed throughout design, implementation, and ongoing oversight^25,26. In practice, this process would unfold as a series of concrete, institutionally embedded steps.

Defining ethical aims and governance mandate

An ethically grounded implementation begins with the clear articulation of institutional ethical goals and a defined governance mandate²⁷. Institutions should specify intended outcomes—such as reducing clinician documentation burden while maintaining note accuracy and protecting patient privacy—and delineate the scope of deployment across specialties, visit types, and care settings. This process should be overseen by a formal, multidisciplinary governance committee comprising clinicians, informaticians, legal and compliance experts, ethics reviewers, and patient representatives, responsible for approving use cases and ensuring continuous oversight of safety, accountability, and equity.

Transparent and meaningful disclosure and consent

Ethical deployment requires disclosure processes that ensure genuine patient understanding and voluntary participation^17,27. Patients should receive clear verbal and written explanations that clinical encounters may be recorded using AI systems for documentation, alongside transparent information about how audio and text data will be stored, accessed, and used. Disclosure must explicitly distinguish acceptable uses, such as clinical documentation and quality improvement, from prohibited secondary uses, including marketing or workforce evaluation without additional consent. Patients should also have an accessible opt-out mechanism documented in the electronic health record, with communication approaches standardized, culturally appropriate, and adapted to varying literacy levels to promote informed decision-making.

Institutional policies and governance structures

A defensible implementation depends on robust institutional policies and governance structures that regulate the use of ambient AI scribes²⁷. Data governance policies should define encryption standards, access controls, retention limits, and restrictions on secondary data use, with independent ethical review required for research applications. Clinician protocols must mandate the review and validation of AI-generated notes, prohibit reliance on such systems for clinical decision-making, and establish clear “stop rules” for sensitive situations or patient objections. These policies should be reinforced by continuous monitoring systems that enable routine audits of documentation quality, identification of errors and bias across demographic groups, and systematic incorporation of stakeholder feedback, ensuring integration into existing regulatory and professional frameworks.

Workflow integration and ongoing adaptation

Sustainable implementation requires the integration of ambient AI scribes into clinical workflows in ways that preserve the clinician–patient relationship⁷. Clinicians should be trained not only in the technical use of these systems but also in communication practices that support transparency, address patient concerns, and maintain professional accountability for clinical records. Structured feedback mechanisms—including surveys, focus groups, and incident reporting—should inform ongoing refinements to consent procedures, data governance policies, and clinical protocols. Institutions must also retain the capacity to scale back or discontinue use if evidence indicates harm, diminished trust, or inequitable outcomes.

Ethical justification and proportional safeguards

These measures are not merely procedural but ethically necessary to address core risks associated with ambient clinical AI, including threats to patient autonomy, potential harm from documentation errors, and justice-related inequities. Because such systems introduce complex data flows and surveillance mechanisms that patients cannot easily scrutinize, and because institutional or commercial incentives may prioritize efficiency over safeguards, strong governance and consent structures are essential¹⁷. Although these requirements may introduce operational complexity and slow implementation, they are proportionate to the potential harms—such as misdiagnosis, biased outcomes, and erosion of trust—and should therefore be regarded as non-negotiable elements of responsible deployment.

Conclusion

Ambient clinical AI represents a significant technological advance with the potential to reduce documentation burdens and improve clinician–patient interaction. However, the ethical risks associated with continuous listening systems cannot be overlooked. Without transparent consent, robust governance, and a commitment to privacy and autonomy, ambient AI risks undermining the trust that is foundational to healthcare.

Listening walls should not become silent intrusions that reshape clinical spaces into environments of invisible surveillance. Instead, they must be deployed in a way that respects human dignity, protects patient rights, and strengthens the relationship between clinicians and the people they serve. Responsible, ethical implementation will determine whether ambient AI becomes an ally in patient-centered care—or a source of erosion of the very values medicine seeks to uphold.

Author contributions

E.K. contributed to conceptualization, literature search, writing—original draft, writing—review and editing, and final approval; J.A. contributed to literature search and writing—original draft; D.S.B. contributed to writing—review and editing, and final approval; C.O.A.B. contributed to writing—review and editing and final approval; E.A. contributed to writing—review and editing and final approval.

Data availability

No datasets were generated or analyzed during the current study.

Competing interests

The authors declare no competing interests.