Algorithmic gaslighting in healthcare: mechanisms, harms, and epistemic injustice in context

Abstract

Algorithmic systems are increasingly woven into healthcare, from triage applications to AI-enabled decision support tools. While such systems promise efficiency and diagnostic accuracy, they also perform a normative function: shaping whose knowledge is treated as credible, which symptoms matter, and how patient testimony is interpreted. This article introduces algorithmic gaslighting to name the patterned dynamic through which testimonial and hermeneutical injustices become recursively stabilized and internalized as patient self-doubt in algorithmically mediated clinical encounters. Building on feminist and disability bioethics accounts of medical gaslighting—where clinicians dismiss or reinterpret patient reports, often in gendered or racialized ways—I extend the analysis to algorithmic contexts. Drawing on Miranda Fricker’s theory of epistemic injustice, I show how interlocking sociotechnical mechanisms—data violence, opacity, automated authority, and feedback loops—can displace patient credibility, constrain interpretive resources, and destabilize patient self-trust. These dynamics recalibrate epistemic agency, rendering patients as data objects rather than recognized knowers. To counter such harms, I propose a normative framework aimed at interrupting gaslighting dynamics in algorithmically mediated care, emphasizing rights to contestation, participatory design, critical audits, and shared responsibility. Recognizing algorithmic gaslighting helps safeguard trust, dialogue, and agency in the digital transformation of healthcare.

Introduction

Algorithmic systems are increasingly embedded in healthcare decision-making, from triaging patients through symptom checkers to guiding clinical diagnoses via AI-enabled decision support tools. While these systems promise greater efficiency, coordination, and diagnostic accuracy, they also perform an under-examined normative function: they shape whose knowledge is treated as credible, which symptoms are considered salient, and how patient testimony should be interpreted. Data-driven systems can reorganize credibility and authority in subtle but consequential ways. In healthcare, where clinical judgment and lived experience are already structured by asymmetries of expertise and power, such reorganization raises urgent ethical questions about epistemic agency and trust.

In response to these developments, I introduce the concept of algorithmic gaslighting to capture how epistemic injustice becomes concretely enacted and stabilized within algorithmically mediated clinical encounters. By algorithmic gaslighting, I identify a structural–relational configuration through which credibility distortions and interpretive marginalization become infrastructurally mediated, recursively reinforced, and experientially internalized in practice.

More precisely, algorithmic gaslighting designates a patterned configuration in which: (i) epistemic authority is redistributed within a triadic doctor–patient–technology structure; (ii) patient testimony and interpretive framing are systematically subordinated to algorithmic outputs; (iii) this subordination is stabilized through routinized reliance and infrastructural feedback; and (iv) the resulting pattern contributes to the internalization of epistemic destabilization, such that patients come to doubt the reliability of their own experiential knowledge.

These conditions obtain independently of malicious intent. The concern is not isolated error or bias, but a patterned reorganization of credibility and interpretive uptake that restructures how epistemic agency functions within digitally mediated clinical practice. This claim concerns intent to harm or deceive a particular patient, which is absent; it does not extend to the prior institutional choice to adopt standardizing, case-based information infrastructures over narrative ones, a choice that is often made deliberately and for reasons explored in Section “Organizational intentionality and the politics of standardization”.

The notion of “medical gaslighting” has emerged in feminist and disability bioethics literature to describe situations in which patients’ reported symptoms are dismissed, minimized, or reinterpreted by healthcare providers—often grounded in prejudicial assumptions tied to gender, race, or chronic illness [8, 11]. Bodies of women, racial minorities, and chronically ill individuals are disproportionately subject to such dismissals, as their accounts are rendered less credible or coherent within dominant clinical narratives [11, 57, 68]. Recent philosophical work has further interpreted gaslighting itself as a form of epistemic injustice, especially when understood structurally rather than as a paradigm case of intentional interpersonal deception. On this broader view, what defines gaslighting is not malicious motive but the destabilization of self-trust and interpretive agency through patterned credibility manipulation.

Building on this foundation, this article extends the analysis from interpersonal medical encounters to algorithmically mediated ones. In contemporary healthcare, algorithmic systems participate in shaping clinical authority, guiding professional judgment and structuring patient participation within broader regimes of epistemic asymmetry and technological paternalism. Even absent malicious intent, such systems may decommission patient epistemic agency by favoring data-driven judgments over embodied, narrative knowledge. Hofmann’s [29] account of technological paternalism underscores that technologies can constrain agency without requiring deliberate domination, reinforcing the view that intention is not a necessary condition for epistemic harm in mediated clinical contexts.

Understood through Miranda Fricker’s framework of epistemic injustice, these developments raise concerns about both testimonial and hermeneutical injustice [24]. Testimonial injustice arises when individuals’ credibility is unfairly diminished; hermeneutical injustice occurs when marginalized groups lack adequate interpretive resources to make sense of their experiences. In healthcare, patients may face testimonial downgrading when their symptom reports are overridden by ostensibly objective system outputs, and hermeneutical marginalization when algorithmic categories fail to capture lived experiences [36, 39]. Algorithmic systems can thus replicate and exacerbate established forms of epistemic injustice.

However, the present analysis goes beyond identifying individual instances of testimonial or hermeneutical injustice in algorithmic settings. The concept of algorithmic gaslighting captures how these injustices become co-produced and stabilized within a triadic clinical configuration. The point is not to multiply taxonomies of injustice, but to illuminate a patterned relational dynamic through which credibility reallocation, interpretive narrowing, and infrastructural mediation converge to destabilize epistemic self-trust. While existing scholarship has examined algorithmic bias, opacity, and machine-learning-induced hermeneutical injustice in healthcare contexts, the recursive and internalizing configuration through which such distortions become experientially sedimented in clinical practice has not been systematically conceptualized.

Why retain the term “gaslighting” rather than referring more generally to algorithmic epistemic injustice? The answer lies in the phenomenological dimension of the harm. Gaslighting captures not merely the occurrence of credibility deficits or interpretive gaps, but the destabilization of self-trust that follows from their patterned repetition. In clinical contexts, patients may come to experience algorithmically mediated misrecognition not as a singular disagreement, but as evidence that their own experiential knowledge is unreliable. The term therefore foregrounds the relational and internalized dimension of infrastructurally mediated epistemic injustice—an aspect that analyses focused solely on bias, transparency, or error risk overlooking.

My thesis is that algorithmic gaslighting articulates a patterned relational configuration through which testimonial and hermeneutical injustices become infrastructurally stabilized and experientially internalized in algorithmically mediated healthcare. Unlike traditional bias or error, this dynamic subtly reconfigures the relational dynamics and co-presence in the triadic configuration where the doctor-patient dyad is accompanied by technological mediation [48, 49, 50]. Algorithmic systems may present themselves as neutral arbiters, while embedding structural misrecognitions that erase or distort patient meaning in ways that are experientially disempowering.

This article makes three contributions to debates on AI ethics and healthcare. First, it introduces algorithmic gaslighting as an interpretive lens for identifying how testimonial and hermeneutical injustices become infrastructurally mediated and experientially internalized in algorithmically mediated care. Second, it identifies specific sociotechnical mechanisms—including data violence, opacity, automated authority, and feedback loops—through which algorithmic mediation can systematically destabilize patient epistemic agency, moving beyond prevailing critiques centered on bias, accuracy, or transparency. Third, it develops a normative framework oriented toward interrupting these gaslighting dynamics, proposing principles such as contestation rights, participatory design, and shared responsibility aimed at restoring credibility recognition, interpretive inclusion, and patient self-trust. Together, these contributions reframe current discussions of AI-related harm in medicine by foregrounding epistemic agency as a core ethical stake in the digital transformation of care.

It is worth stating directly, at the outset, what algorithmic mediation adds to gaslighting dynamics already documented in interpersonal clinical encounters. The argument is not simply that algorithms make dismissive clinical judgments occur faster or more often. Three differences are structural rather than merely quantitative. First, algorithmic mediation converts a dyadic doctor–patient relation into a triadic one, introducing a third locus of authority that is answerable to neither party in the way a clinician is answerable to professional and institutional norms. Second, opacity forecloses the kind of reasoning-with that even a dismissive clinician still permits: a patient can, in principle, argue with a person’s stated reasoning, but cannot argue with a score whose derivation is inaccessible to both patient and clinician alike. Third, because algorithmic outputs are infrastructural rather than episodic, the resulting credibility displacement is stabilized and reproduced across cases rather than confined to a single encounter. These three differences are developed at length in Section “From human to algorithmic: inheriting and amplifying injustice” and Section “Integrating testimonial and hermeneutical injustice in algorithmic contexts”; they are previewed here because they bear directly on why the term ‘algorithmic,’ and not merely ‘gaslighting,’ is doing conceptual work in this article.

The paper proceeds as follows. Section “Conceptual foundations” develops the conceptual foundations of algorithmic gaslighting by tracing the shift from interpersonal medical gaslighting to its algorithmically mediated form, situating the analysis within feminist and disability bioethics, postphenomenology, and Fricker’s framework of epistemic injustice. Section “Mechanisms and harms of algorithmic gaslighting in healthcare” then identifies the sociotechnical mechanisms—data violence, opacity, automated authority, and feedback loops—through which algorithmic gaslighting is enacted in triadic doctor–patient–technology configurations, and analyzes the patterned epistemic harms they generate. Section “A normative framework for interrupting algorithmic gaslighting in healthcare” advances a normative framework aimed at interrupting these dynamics, articulating interventions at the level of individual agency, system design, and institutional governance. The Conclusion reflects on the broader implications of recognizing algorithmic gaslighting as a structural configuration of epistemic injustice in digitally mediated healthcare.

By drawing attention to these silent yet pervasive harms, this article aims to foreground patient epistemic agency as a constitutive ethical concern in the deployment of AI in healthcare. Recognizing and addressing algorithmic gaslighting is essential not only for fairness, but for preserving the conditions under which medicine can remain trustworthy, dialogical and responsive to those it serves.

Conceptual foundations

From medical gaslighting to the algorithmic form

The term medical gaslighting has gained traction in feminist and patient-advocacy discourse to characterize clinical situations in which healthcare providers dismiss, trivialize, or misattribute patients’ symptoms—often attributing them to psychological causes or otherwise negating lived experience. This phenomenon disproportionately affects women, racial minorities, and those with chronic or contested conditions [25, 35], leading to diagnostic delays, retraumatization, and erosion of trust. Feminist bioethics critiques this dynamic through its broader interrogation of androcentrism, power imbalances, and the marginalization of non-dominant voices in healthcare [25, 35, 63]. Importantly, systematic review evidence underscores how medical gaslighting engenders anxiety, isolation, and entrenched gender bias in clinical interactions [35].

Beyond the individual level, systemic and institutional forces contribute to medical gaslighting. Marginalized groups, such as Black women during pregnancy, report encounters where stereotypes and epistemic dismissal compound inequities [11]. Feminist bioethicists argue these dynamics are not incidental but rooted in oppressive epistemic structures that prioritize biomedical expertise over embodied knowledge, reinforcing patients’ disenfranchisement [23, 62, 67].

A recurrent critique of the term “gaslighting” in medical contexts is that it misapplies a concept originally tied to interpersonal deception. In its paradigmatic sense, gaslighting refers to deliberate manipulation, where one agent seeks to destabilize another’s self-trust for strategic or malicious ends. By contrast, clinicians who dismiss or reinterpret patients’ accounts often do not act with such intention; their responses are shaped by implicit bias, institutional routines, and systemic epistemic hierarchies. On this view, medical gaslighting risks being a misleading metaphor, since intent is absent [26].

However, feminist and disability bioethics have already argued for a broadened understanding of gaslighting that shifts attention from ill-intent to structural conditions [10]. What unites gaslighting practices is not the subjective aim of the perpetrator but the relational outcome: the patient is rendered doubtful of their perceptions, deprived of credibility, and alienated from their interpretive resources needed to make sense of their experience. It is this experiential dimension—the destabilization of epistemic self-trust through repeated credibility dismissal —rather than intentional deception, that defines the phenomenon.

Philosophical analyses have extended this insight by theorizing gaslighting as a form of epistemic injustice in its own right. Abramson, [1] characterizes gaslighting as a manipulation that destabilizes a person’s self-trust as a rational agent, targeting the victim’s capacity to take their own perceptions and judgments as authoritative. McKinnon [32] argues that gaslighting constitutes a species of testimonial injustice while additionally involving a betrayal of moral trust: the target is not merely disbelieved but wronged by the very person or authority to whom they turned for epistemic recognition. Bailey, [7] frames the relationship between gaslighting and epistemic injustice as a productive site of inquiry, noting that gaslighting’s distinctive harms — particularly the erosion of self-trust — exceed what Fricker’s categories of testimonial and hermeneutical injustice alone capture. Levy, [38] extends this analysis to medical contexts, explicitly theorizing medical gaslighting through both testimonial and hermeneutical injustice and emphasizing its cumulative, iterative, self-trust-eroding character: through repeated clinical dismissal, patients come to doubt the reliability of their own bodily perceptions. The present analysis inherits rather than re-establishes this epistemic injustice framing—retaining the term gaslighting as a deliberate analytical choice—and extends it to algorithmically mediated contexts, where credibility destabilization becomes infrastructurally stabilized and recursively reinforced within triadic clinical configurations.

This conceptual shift becomes even more crucial in the algorithmic domain. Algorithms, by design, lack intentions; they cannot plot or deceive in the interpersonal sense. If intentionality were taken as a necessary condition, the very possibility of “algorithmic gaslighting” would collapse. This is a claim about the algorithm as artifact, not about the organizations that select, procure, and deploy it; Section “Organizational intentionality and the politics of standardization” addresses the latter directly. Yet algorithms can clearly replicate—and in some cases intensify—the disorienting dynamics of gaslighting: they override patient testimony with ostensibly neutral authority, reshape symptom narratives into rigid categories, and compel patients to mistrust their embodied experiences. Their operation demonstrates why focusing on intent obscures the deeper normative point. What matters is not whether a clinician or system “means” to gaslight, but whether the relational effect leaves patients epistemically disoriented, undermined, and estranged from their own agency. The strength of the concept lies precisely in naming these relational harms as gaslighting, regardless of intentionality.

While algorithms cannot harbor motives in the human sense, postphenomenological accounts of technological mediation complicate the critique further. Technologies, as Don Ihde [31] and Peter-Paul Verbeek [72] argue, possess a kind of technological intentionality: they direct attention, shape interpretive horizons, and invite specific responses. A diagnostic AI that privileges statistical correlations over narrative reports “intends” the patient’s body and symptoms in a particular way, highlighting what is legible to the model and sidelining what is not. From this perspective, algorithmic gaslighting is not an oxymoron but a paradigmatic case: the system embodies and enacts intentional structures that destabilize patient self-trust and interpretive agency. Even if one insists on intentionality as a definitional feature of gaslighting, postphenomenological accounts allow algorithms to be understood as participating in structured intentional relations through their mediating role in clinical relations. Seen in this light, algorithmic gaslighting should therefore be understood less as a question of deceptive motive and more as a relational enactment through which epistemic injustices become operationalized in clinical contexts. This brings us directly to Fricker’s framework, which clarifies how testimonial and hermeneutical dynamics structure these harms.

Epistemic injustice in healthcare

Miranda Fricker’s framework of epistemic injustice is foundational for understanding how knowledge-related harms operate, especially in the medical context. Testimonial injustice arises when a speaker—here, a patient—is accorded diminished credibility on the basis of identity or perceived irrationality [24]. In healthcare, patients may be presumed cognitively unreliable or overly emotional, undermining their testimonial authority [14]. Hermeneutical injustice, meanwhile, occurs when individuals cannot make sense of or articulate their experiences due to collective interpretive shortcomings; e.g., pre-medical awareness of concepts such as sexual harassment facilitated its recognition and contestation [24]. In healthcare, patients may lack adequate conceptual vocabulary or interpretive frameworks for emerging or complex symptoms—such as those related to long COVID or chronic pain [5].

Carel and Kidd, [14] directly applies Fricker’s insights to healthcare, arguing that ill persons are especially vulnerable to both forms of epistemic injustice: they face testimonial downgrading and structural obstacles in finding interpretive resources to understand their embodied experience, rooted in epistemic privilege held by clinicians and medical institutions [14]. Intersectional analyses—such as those considering race, gender, and age—reveal that patients with overlapping marginalized identities face compounded testimonial injustice, further undermining equitable care [3, 4].

This uptake of Fricker’s framework in healthcare has not gone unchallenged. Nielsen et al. [46] argue that the field lacks robust empiricalevidence and a clear application criterion for identifying testimonial injustice in clinical cases, cautioning that many core claims remain untestedhypotheses. In response, Færge et al. [21] contend that demanding large-scale quantification before acknowledging epistemic injustice risksreproducing the very exclusionary hierarchies the concept is meant to expose, while Floris [22] argues that in some domains—such as women’sdiagnostic delay—existing medical data already provide sufficient evidence of epistemic injustice without awaiting standardized metrics. Hummel [30] extends this evidentiary debate specifically to AI-mediated care, maintaining that claims of AI-induced epistemic injustice require demonstrableevidence both of its presence and of its normative significance relative to competing considerations such as diagnostic accuracy.

From human to algorithmic: inheriting and amplifying injustice

As healthcare becomes increasingly mediated by algorithmic systems, established forms of epistemic injustices risk being inherited, obscured, or amplified. Recent scholarship has shown that machine learning systems do not merely introduce new technical errors into clinical decision making but reshape the epistemic conditions under which credibility, interpretation, and understanding are negotiated.

Milano and Prunkl [42] describe this shift in terms of epistemic fragmentation: algorithmic personalization isolates individuals within data-defined categories, making collective meaning-making more difficult and weakening shared interpretive resources. In such contexts, individuals may lack the epistemic footholds necessary to situate their experiences within a broader communal framework.

More concretely, Pozzi [52] demonstrates how medical machine learning systems can generate testimonial injustice by functioning as markers of trustworthiness. In her analysis of opioid risk scoring platforms (PDMPs), algorithmic outputs come to mediate clinicians’ credibility assessments, such that patients’ testimony is deflated not on epistemically valid grounds but because risk scores are treated as authoritative indicators. Similarly, empirical analyses indicate that predictive systems can overrule or discount patient-reported symptoms when model outputs are perceived as more “objective” or reliable [34]. In such cases, the injustice does not lie in isolated prejudice but in how algorithmic mediation reorganizes credibility hierarchies within clinical encounters.

Beyond testimonial distortion, ML systems may also contribute to hermeneutical injustice. Pozzi [53] argues that certain systems engage in what she calls automated hermeneutical appropriation: algorithmic platforms establish interpretive categories—such as “risk of misuse”—without meaningful human oversight, thereby shaping shared hermeneutical resources while limiting patients’ ability to contest or reinterpret these imposed meanings. In broader AI contexts, related concerns emerge under the label of generative hermeneutical erasure, where large generative models systematically marginalize non-Western epistemic frameworks and erase conceptual particularities of underrepresented knowledge traditions [45]. Although these analyses operate at different scales, they converge in identifying algorithmic systems as participants in the production—and sometimes distortion—of shared interpretive resources.

Relatedly, Pozzi and Durán [55] argue that epistemic opacity in medical ML is not simply an informational limitation but can produce forms of epistemic objectification (see also [54]). When explanatory structures are inaccessible and ethically insulated, patients risk being positioned as passive objects of algorithmic assessment rather than participants in shared epistemic practices. Ethical features of the clinical situation, they argue, should play a regulatory role in epistemic evaluation—a point that underscores how deeply intertwined epistemology and ethics are in algorithmically mediated care.

Taken together, these strands of scholarship converge on a shared insight, recently synthesized by Faissner et al. [20]: algorithmic systems can reshape epistemic standing not only through discrete credibility errors but through structural features of technological mediation. They demonstrate that testimonial and hermeneutical injustices can be produced or exacerbated when algorithmic outputs acquire institutional authority, when opacity forecloses meaningful recourse, and when interpretive categories are computationally stabilized without adequate deliberative engagement.

A foundational structural account of these dynamics is provided by Symons and Alvarado [66], who argue that data science technologies can generate epistemic injustices understood as harms done to persons specifically in their capacity as knowers. They show that such harms need not arise from explicit prejudice; rather, they can emerge from the institutional authority and opacity characteristic of algorithmic systems. When algorithmic determinations are closed to epistemic recourse—difficult to interrogate, contest, or reinterpret—individuals may be rendered epistemically subordinate to outputs whose justificatory structure remains inaccessible. Although their analysis focuses primarily on legal and institutional hierarchies, the same structural dynamic is evident in clinical settings, where algorithmic outputs can mediate credibility judgments between doctors and patients.

Building on these structural accounts, the concept of algorithmic gaslighting identifies a further pattern within algorithmically mediated healthcare. First, while existing analyses convincingly demonstrate how testimonial and hermeneutical injustices occur in specific cases, algorithmic gaslighting names the process by which such injustices become recursively stabilized through routinized reliance on algorithmic outputs and feedback loops within institutional practice. Second, beyond external foreclosure of contestation, repeated credibility displacement can produce internal effects: patients may come to doubt the reliability of their own experiential knowledge, not merely because they are overruled, but because algorithmic authority is sedimented as epistemically superior. Third, whereas much of the existing literature examines dyadic relations between subordinated individuals and institutional authorities, clinical contexts introduce a triadic configuration—doctor, patient, and algorithm—in which credibility assessments are mediated rather than directly imposed. In this triadic arrangement, epistemic authority is reallocated indirectly through technological infrastructure, allowing destabilizing dynamics to persist even in the absence of malicious intent.

Organizational intentionality and the politics of standardization

The claim that algorithmic gaslighting does not require malicious intent should not be read as a claim that no one intends anything in this process. Algorithms do not select themselves; healthcare organizations choose to adopt them, and that choice is frequently made with an explicit institutional preference for standardized, case-based information over narrative, embodied accounts. Kiviat [37] argues that rendering people as cases — discrete entities defined by regularized, atemporal attributes — is a distinct cultural and informational form from understanding people as actors in the unfolding narratives of their lives, and that this choice of form carries its own moral consequences: casing enables comparative justice (treating like cases alike) while displacing the noncomparative judgments that narrative accounts support. When a hospital system adopts a triage algorithm, a risk score, or a structured symptom-intake form in place of open-ended clinical narrative, it is not a neutral technical upgrade but an intentional reallocation of which informational form will govern clinical judgment.

This administrative preference for legibility has a longer history than healthcare AI. Scott [61] shows that state and institutional actors have long favored standardized, simplified representations of populations — the legible map over the lived landscape — because legibility is what allows large organizations to govern, audit, and scale their operations; what is illegible to the administrative apparatus is, in a precise sense, not governable by it. Bowker and Star [12] extend this insight to classification infrastructures specifically, showing that the categories embedded in institutional information systems are themselves the product of contested organizational choices about what counts as a relevant distinction, choices that then become invisible once naturalized into routine practice. Healthcare organizations that adopt algorithmic risk scores, structured intake forms, or automated triage tools are participating in exactly this kind of legibility project: they are intentionally exchanging the richness and contestability of narrative testimony for the auditability and portability of standardized case data, typically for reasons of efficiency, liability management, or scale.

This reframing does not collapse the distinction on which the article’s central thesis depends. The thesis is that algorithmic gaslighting does not require intent to deceive or harm any particular patient — and that claim survives intact. What must be added is that the infrastructural conditions under which such deception-free harm becomes possible are not themselves accidental. An organization’s decision to privilege standardized case data over patient narrative is an intentional act of institutional design, even when no one downstream intends to disbelieve or destabilize any individual patient. Locating intent at this upstream, organizational level — rather than at the level of the algorithm or the individual clinician — clarifies where responsibility properly attaches (a point developed further in Section “A normative framework for interrupting algorithmic gaslighting in healthcare’s” discussion of shared responsibility) without requiring the article to abandon its claim that the phenomenon at the point of care is gaslighting without a gaslighter.

When testimonial authority is displaced and interpretive resources are constrained in ways that undermine patients’ confidence in their own experiential knowledge, these dynamics align with what this article characterizes as algorithmic gaslighting. The following section therefore turns to the sociotechnical mechanisms through which this recursive destabilization is enacted in clinical practice.

Mechanisms and harms of algorithmic gaslighting in healthcare

Algorithmic gaslighting is enacted within clinical encounters where algorithmic outputs mediate the exchange between doctor and patient through interlocking sociotechnical mechanisms that systematically undermine patient epistemic agency. These mechanisms (see Table 1) —data violence, opacity, automated authority, and feedback loops—are not simply technical imperfections or instances of bias. In brief, and before turning to each in detail: data violence concerns whose bodies are legible to the system at all, since training data that underrepresents a group renders that group’s symptoms statistically invisible; opacity concerns whether a given output can be interrogated once produced; automated authority concerns whether the output, once produced, is treated as more credible than the patient’s or clinician’s own judgment; and feedback loops concern how repeated exposure to the first three dynamics reshapes what patients and clinicians come to say and expect in future encounters. The remainder of this section takes up each mechanism in turn. They constitute epistemic structures that reshape credibility, interpretive resources, and trust in healthcare contexts. While these are the mechanisms I have focused on in this inquiry, I do not claim they represent the only ways in which algorithmic gaslighting can manifest. Their operation produces harms that extend beyond misdiagnosis or error, generating patterned epistemic injuries —erosion of self-trust, collapse of institutional trust, epistemic objectification, and the recalibration of agency— through which testimonial and hermeneutical injustices become experientially stabilized. This section traces how these mechanisms produce relational pathways of epistemic destabilization.

Table 1.

Mechanisms of algorithmic gaslighting

Mechanism (Section)	What it is	Injustice produced	Patterned harm	Illustrative case
Data violence (Section “Data violence - epistemic erasure and objectification”)	Training data and devices statistically underrepresent certain bodies, making their symptoms illegible to the system	Testimonial (symptoms discounted) + upstream hermeneutical (categories built without these patients in view)	Patient’s bodily knowledge doesn’t “register” — internalized as unreliability of their own perception	Pulse oximeters overestimating blood oxygen in Black patients [64]
Opacity (Section “Opacity - undermining self-trust and contestability”)	The reasoning behind an algorithmic output cannot be interrogated, even by treating clinicians	Testimonial (patient cannot contest the score that overrides their report)	No “reasoning-with” is possible, unlike with a dismissive clinician — the patient has nothing to argue against	Epic Sepsis Model’s inaccessible proprietary scoring logic [75]
Automated authority (Section “Automated authority - collapse of institutional trust”)	Algorithmic outputs are treated as objectively superior to human clinical judgment	Testimonial (narrative knowledge loses out to “neutral” score)	Patient is pressured to doubt symptoms when “the algorithm says otherwise”; institutional trust collapses when authority outruns evidence	IBM Watson for Oncology’s unsafe recommendations despite marketed authority [59]
Feedback loops (Section “Feedback loops - distortion of self-narration and epistemic agency”)	Patients and clinicians adapt reporting practices to fit algorithmic categories, which become further entrenched	Hermeneutical (interpretive resources narrow to what is machine-legible)	Patient’s self-narration is reshaped from the inside — a Hacking-style looping effect, but looping through retraining rather than conscious reclassification	EHR language patterns recurring across encounters and feeding future model calibration [65]

Data violence → epistemic erasure and objectification

The first mechanism through which algorithmic systems enact gaslighting is what I refer to as data violence. Drawing on Mollema’s [45] account of generative hermeneutical erasure I use this term as a shorthand to designate the upstream infrastructural suppression and distortion of marginalized epistemologies within training datasets. Training datasets in healthcare often privilege white, male, and socioeconomically advantaged populations, leading to the statistical underrepresentation of women, racial minorities, and people with chronic or contested conditions [27]. The result is not just uneven accuracy but epistemic invisibility: symptom patterns common to underrepresented groups are rendered algorithmically illegible. This means that patients whose embodied experiences are absent from dominant datasets become unrecognizable to algorithmic mediation. For instance, chest pain in women often presents differently than in men, yet cardiovascular datasets disproportionately encode male symptom profiles [16]. Algorithms trained on such data reproduce this epistemic skew, misclassifying women’s symptoms or failing to detect them altogether. A parallel dynamic affects race: pulse oximeters, calibrated predominantly on lighter-skinned patients, systematically overestimate blood oxygen saturation in Black patients, producing occult hypoxemia that is not reflected in the device reading [64]. Because this device-level measurement error propagates into downstream triage and treatment algorithms that treat oximetry readings as ground truth, it exemplifies how data violence operates not only through the absence of a group from training data but through devices and pipelines calibrated on an unrepresentative population and then deployed universally. For example, Rickman [58] finds that summaries produced by certain large language models in long-term care settings underemphasize women’s physical and mental health issues, or use more indirect or euphemistic language for them, compared to summaries generated for men.

The epistemic consequence is twofold. First, patients excluded from training corpora experience epistemic erasure: their bodily knowledge does not register within the algorithmic frame. Second, this erasure contributes to epistemic objectification. Instead of being treated as knowers, patients become reduced to data objects that either “fit” the system’s categories or are disregarded Erbay [19]. Fricker [24] defines objectification as treating individuals not as epistemic agents but as passive sources of information; Medina [41] expands this to include stripping interpretive authority. Algorithmic systems enact precisely this move: extracting data points while denying patients’ capacity to shape the meaning of their narratives.

This dynamic also corrodes self-trust. As Carel and Kidd [14] argue, ill persons are particularly vulnerable to testimonial injustice, and when their accounts are repeatedly invalidated by algorithms, they may begin to doubt the reliability of their perceptions. Patients whose symptoms are misrecognized or erased by AI outputs internalize the message that their lived experiences are epistemically unreliable. Over time, this undermines confidence in their own bodily awareness, leading to disengagement from clinical dialogue and even self-silencing.

Opacity → undermining self-trust and contestability

Opacity constitutes the second mechanism through which algorithmic gaslighting is enacted. The “black-box” character of many machine learning systems is not merely an informational inconvenience but an epistemically consequential feature that reconfigures relations of authority within clinical encounters. Symons and Alvarado [66] argue that data science systems can generate epistemic injustice in part because they are frequently closed to epistemic recourse. Unlike the testimony of a prejudiced witness—which can be revisited, cross-examined, and challenged—algorithmic determinations are often difficult to interrogate, even by those who deploy them. When justificatory pathways are inaccessible, individuals affected by such determinations lack meaningful avenues to contest or reinterpret them.

The Epic Sepsis Model, a proprietary early-warning algorithm deployed at hundreds of U.S. hospitals, illustrates the stakes of this opacity concretely: an external validation study found the model performed far worse than its developer’s internal claims, correctly identifying only a third of sepsis cases while generating frequent false alerts, yet neither treating clinicians nor patients had access to the proprietary scoring logic needed to understand or contest any individual alert [75].

In healthcare contexts, this structural opacity has context-sensitive implications. A patient whose symptoms are misclassified by a predictive model may have no legible way to challenge the inference, not because appeal is formally prohibited, but because the inferential process itself is epistemically opaque. In such contexts, algorithmic opacity produces what has been described as a “reliability gap,” wherein neither patients nor physicians can independently assess the justificatory structure of system outputs, thereby intensifying epistemic asymmetries within the triadic clinical arrangement [13]. Even when explanation tools are provided, they may be technical, abstract, or detached from the patient’s lived experience [2, 40]. As Pozzi and Durán [55] note, opacity can position patients as objects of assessment rather than participants in epistemic exchange, thereby altering the normative structure of clinical dialogue.

Opacity also interacts with testimonial dynamics. As Pozzi [53] argues, machine learning systems in healthcare can function as markers of trustworthiness, mediating clinicians’ credibility judgments. In such cases, patient testimony may be deflated not because the system itself “disbelieves” the patient, but because algorithmic outputs are treated as epistemically privileged indicators. Opacity intensifies this dynamic: when the grounds of the output cannot be meaningfully scrutinized, its authority becomes harder to resist. What appears as neutral computational objectivity can thereby displace both patient narrative and clinician judgment. Empirical work suggests that reliance on predictive scores can contribute to communication breakdowns and credibility deficits in practice [34], reinforcing epistemic asymmetries within care relationships.

This foreclosure of epistemic recourse bears directly on contestability. Symons and Alvarado [66] emphasize that injustice arises not only from biased outcomes but from the diminished capacity of individuals to challenge algorithmic determinations, a concern echoed in Vaassen’s [70] analysis of opacity as a threat to personal autonomy. Similarly, Edwards and Veale [18] argue that rights to explanation should be understood as enabling better decisions—providing grounds upon which automated judgments may be interrogated and revised. When opacity renders such interrogation practically unavailable, contestation is not simply difficult but structurally constrained. Patients may find themselves unable to articulate why a system’s output fails to capture their experience, because the criteria by which that output was generated remain obscure.

Such structurally foreclosed contestability intensifies epistemic and power asymmetries characteristic of triadic clinical arrangements, where algorithmic systems mediate credibility judgments between doctor and patient [48]. Over time, this structural non-contestability can undermine self-trust. Repeated experiences in which algorithmic outputs override or discount experiential reports may lead patients to internalize credibility deficits. A person reporting persistent pain, for instance, may begin to doubt the reliability of their own sensations when predictive systems repeatedly fail to register them as clinically significant. This dynamic resembles what Carel and Kidd [14] describe as testimonial downgrading, yet in algorithmically mediated contexts the downgrading is infrastructurally mediated rather than solely interpersonal. The resulting erosion of epistemic confidence does not stem from explicit accusation or intentional deception but from the sedimented authority of opaque computational systems. In this way, opacity contributes directly to gaslighting dynamics by destabilizing patients’ confidence in their own experiential knowledge while reinforcing asymmetries between algorithm, clinician, and patient.

Automated authority → collapse of institutional trust

Algorithmic systems do not operate as neutral tools; rather, they are often framed within clinical and institutional contexts as objective and superior arbiters of truth. This framing confers what may be described as automated authority, whereby outputs are treated as more credible than human judgment. The perception of neutrality, coupled with opacity, fosters what Langdon Winner calls “technological somnambulism”: the uncritical acceptance of technological artifacts as inherently beneficial [74].

In clinical settings, this leads both patients and clinicians to defer to machine outputs, even against their better judgment. When the algorithm “says otherwise,” patients are pressured to doubt their symptom narratives, while clinicians may hesitate to override the supposedly objective recommendation. Automated authority thus amplifies testimonial injustice, privileging statistical authority over narrative knowledge. Such credibility reordering contributes to gaslighting dynamics by positioning algorithmic outputs as epistemically dominant while rendering patient narratives increasingly defeasible, thereby intensifying experiential self-doubt.

The broader consequence is a collapse of institutional trust. Trust in healthcare institutions is cumulative and relational, grounded in perceptions of competence, benevolence, and alignment of values [6, 51]. When algorithms reproduce systemic inequities—such as risk prediction tools underestimating illness severity in Black patients due to cost-based proxies [47]—patients recognize that the system is not impartial but structurally biased. Automated authority can also outrun the evidence supporting it: internal IBM documents later revealed that Watson for Oncology, marketed to hospitals worldwide as a decision-support tool trained on real patient outcomes, had in fact been trained substantially on synthetic, hypothetical cases and generated ‘multiple examples of unsafe and incorrect treatment recommendations’ even as the product was being promoted as authoritative [59]. The episode illustrates that automated authority can be extended to a system independently of any demonstrated reliability, and that this authority — once granted — is not easily withdrawn even when the underlying performance is shown to be poor. The veneer of neutrality masks deep-seated injustices, leading to disillusionment not only with a particular encounter but with the institution as a whole.

This collapse of trust is compounded by the accountability gap inherent in AI-mediated care [60]. Responsibility is dispersed across developers, vendors, clinicians, and administrators, leaving patients with no clear avenue for redress. The recognition that the system is both untrustworthy and unaccountable undermines the moral contract between patient and institution, entrenching skepticism and withdrawal. In this way, automated authority extends gaslighting dynamics beyond individual encounters, embedding epistemic destabilization within institutional structures and transforming mistrust into a systemic relational condition rather than an episodic response.

Feedback loops → distortion of self-narration and epistemic agency

The final mechanism, feedback loops, describes how algorithmic categories reshape not only diagnoses but also patient self-understanding. As patients and clinicians adapt their reporting practices to align with model expectations, symptom narratives become tailored to algorithmic categories [52]. Over time, patients may internalize these categories, reframing their experiences to fit the system’s interpretive grid.

This process distorts self-narration. Patients no longer describe their symptoms in ways that reflect their lived realities but in ways designed to be legible to the machine. Hermeneutical injustice arises as the interpretive resources available to patients are narrowed to those sanctioned by algorithmic logics [24, 36]. A concrete instance of this narrowing is visible in electronic health record documentation itself: language used to describe patients — including stigmatizing or dismissive phrasing — recurs across encounters and providers once it enters the record, shaping how subsequent clinicians interpret the same patient and, where such documentation feeds model training or evaluation, how future algorithmic outputs are calibrated ([65]; discussed further in Section “Cumulative and intersectional impacts”). The epistemic harm is not only exclusion but transformation: patients’ interpretive autonomy is eroded, and their narratives reshaped to conform to algorithmically legible categories. This progressive reshaping of self-interpretation exemplifies gaslighting dynamics by relocating epistemic authority outside the patient while subtly reorganizing how experience itself is articulated and understood.

Such feedback loops are self-reinforcing. As more patients adapt their testimonies to algorithmic categories, these categories become further entrenched, narrowing the conceptual space for alternative expressions. Mills [43] describes this dynamic as an “epistemology of ignorance,” wherein structures of knowledge perpetuate their own blind spots. In the clinical context, patients unknowingly perpetuate the very frameworks that erase their experiences, thereby recalibrating their epistemic agency to function within algorithmically defined parameters. This dynamic closely resembles what Hacking [28] calls the ‘looping effect’ of human kinds: a classification does not merely describe the people it sorts but changes them, as classified individuals adjust their self-understanding and behavior to fit the category, which is then revised in light of that very adjustment, moving the target of classification over time. Symptom categories embedded in algorithmic triage or risk-scoring tools function analogously to the clinical classifications Hacking examines: patients learn, often through repeated clinical interaction, which descriptions of their experience will register as legible ‘cases’ within the system, and they adjust their self-reporting accordingly. Unlike Hacking’s classifications, which loop through conscious human interpretation and reclassification, algorithmic categories loop through retraining and infrastructural reuse — the adjusted reporting becomes new training or documentation data, further entrenching the category it was shaped to fit. The loop is therefore not merely descriptive drift but a mechanism through which patients’ own adaptive responses to misrecognition end up reinforcing the categories responsible for that misrecognition. Through this recursive stabilization, feedback loops illustrate the pattern-level dimension of algorithmic gaslighting: epistemic destabilization is no longer episodic but becomes sedimented within interpretive practice, reshaping agency over time.

Integrating testimonial and hermeneutical injustice in algorithmic contexts

The preceding analysis raises a clarificatory question: do the mechanisms described here merely instantiate already recognized forms of testimonial and hermeneutical injustice? If so, what conceptual work does the notion of algorithmic gaslighting perform?

The account developed in this article does not introduce a third subtype of epistemic injustice. Rather, algorithmic gaslighting designates a structural configuration in which testimonial and hermeneutical injustices become infrastructurally integrated, mutually reinforcing, and experientially internalized within algorithmically mediated clinical relations. While testimonial injustice concerns credibility deficits and hermeneutical injustice concerns interpretive gaps, algorithmic gaslighting identifies how these distortions co-function within a technologically stabilized epistemic environment. Three structural features distinguish this configuration.

First, credibility reallocation occurs within technologically mediated clinical arrangements in which algorithmic outputs participate in shaping authority. Rather than operating solely through direct interpersonal judgment, credibility assessments are filtered through computational categories, predictive scores, and classification systems. Epistemic authority is therefore mediated through infrastructure rather than exercised exclusively between persons.

Second, these dynamics become stabilized through routinized reliance and feedback loops. Algorithmic systems embed categories and credibility hierarchies into workflows, documentation practices, and institutional routines. As patients and clinicians adjust their practices to align with system outputs, these hierarchies become sedimented within the epistemic infrastructure of care. What might otherwise remain episodic misrecognition becomes a normalized feature of clinical interaction.

Third, and most importantly, such stabilization can reshape epistemic self-relation. Repeated experiences of credibility displacement and interpretive narrowing may erode self-trust. As Levy [38] argues in the interpersonal case, medical gaslighting can cumulatively undermine confidence in one’s own perceptions. In algorithmically mediated contexts, similar self-trust–eroding effects arise not only through repeated interpersonal dismissal but through the infrastructural reinforcement of algorithmic authority. Over time, patients may recalibrate their epistemic self-conception in light of persistent computational mediation.

What is conceptually distinctive, then, is not the identification of new mechanisms of injustice, but the articulation of a patterned structural transformation in how established injustices co-function within digital clinical systems. Existing scholarship analyzes algorithmic testimonial injustice, automated hermeneutical injustice, opacity-induced distortions, and epistemic objectification largely as discrete harms. The present account isolates their infrastructural integration into a self-reinforcing epistemic environment in which credibility displacement, interpretive narrowing, and automated authority operate together as a stabilized relational configuration.

The distinctive harm of algorithmic gaslighting therefore lies at the level of epistemic subject-formation. Patients are not only disbelieved or misunderstood in particular instances; they are gradually repositioned within a system that structurally recalibrates the conditions under which their testimony is received and their experiences are interpreted. It is this patterned transformation—where testimonial and hermeneutical distortions become infrastructurally sedimented and experientially internalized—that warrants the integrative designation “algorithmic gaslighting.”

Cumulative and intersectional impacts

The cumulative effects of these mechanisms are magnified for patients with intersecting marginalized identities. A Black woman with a disability, for example, may encounter compounded testimonial injustice (credibility deficits tied to race and gender) and hermeneutical injustice (lack of adequate frameworks for disability experience). When algorithms trained on datasets underrepresenting both her racial group and her condition mediate her care, these credibility deficits are not merely repeated but infrastructurally reinforced. Misrecognition may shift from episodic interpersonal failure to a pattern embedded within statistical modeling and clinical workflow.

Studies of negative patient descriptors in electronic health records demonstrate how clinician language can create persistent patterns of stigmatization that are reproduced across encounters and providers [65]. When such documentation practices become incorporated into data infrastructures or used in the training and evaluation of predictive systems, the risk of amplifying existing patterns of misrecognition increases. In this way, algorithmic mediation can interact with already biased documentation practices to stabilize rather than merely reproduce intersectional credibility deficits.

Importantly, algorithmic mediation transforms intersectional injustice from a contingent interpersonal phenomenon into a stabilized feature of epistemic infrastructure. Because datasets, predictive categories, and feedback loops recursively shape future clinical encounters, misrecognition can accumulate across time and cases. The injustice is no longer confined to individual credibility judgments but becomes encoded in system outputs and institutional routines.

These harms also scale beyond the individual. Communities subjected to repeated algorithmic misrecognition may develop collective mistrust of healthcare institutions, reinforcing disparities in access, diagnosis, and treatment [9]. In this way, algorithmic gaslighting does not merely reproduce existing inequities; it can amplify and sediment them through infrastructural mediation.

Taken together, data violence, opacity, automated authority, and feedback loops generate a pattern in which epistemic injustice becomes both cumulative and systematically stabilized. They erase marginalized experiences, undermine patient and institutional trust, objectify patients within their own care, and recalibrate epistemic agency itself. Marginalized experiences are not simply discounted but progressively rendered illegible within algorithmically structured care environments. Algorithmic gaslighting thus designates the process through which intersectional epistemic injustice becomes infrastructurally mediated and recursively reinforced. Addressing it therefore requires structural redress aimed not only at improving fairness metrics but at preventing the sedimentation of misrecognition within the epistemic architecture of healthcare.

A normative framework for interrupting algorithmic gaslighting in healthcare

If algorithmic gaslighting is understood as the structural configuration through which testimonial and hermeneutical injustice become recursively stabilized within algorithmically mediated care, then responding to it requires more than generic remedies for epistemic injustice. The challenge is not only to correct credibility deficits or expand interpretive resources in isolation, but to interrupt the relational configuration through which algorithmic authority recalibrates epistemic agency over time. Because the harm operates through infrastructural mediation rather than isolated interpersonal misjudgment, remedies must address not only epistemic recognition in encounters but the design and governance of the systems that shape those encounters.

Standard fairness or transparency interventions address distributive and procedural concerns, but they do not necessarily confront the triadic authority structure, recursive feedback dynamics, and internalization of self-doubt identified above. A normative response must therefore target the relational architecture of algorithmically mediated encounters.

Such a framework can be articulated along three interrelated dimensions. The first dimension concerns the restoration of epistemic agency at the level of individual encounters, where patients must be able to contest algorithmic outputs and receive explanations that support shared understanding. The second dimension concerns the structural embedding of epistemic justice within the design and operation of healthcare technologies, recognizing that hermeneutical resources are shaped by collective processes rather than individual transactions. The third dimension concerns shared responsibility, which requires rethinking accountability across the distributed network of developers, clinicians, institutions, and regulators. Only by addressing all three dimensions together can we interrupt the structural dynamics through which testimonial and hermeneutical injustices become stabilized in algorithmically mediated care.

At the level of individual agency, the right to contestation occupies a central place. Contestation should not be understood narrowly as appealing a decision or requesting a second opinion, but rather as a broader epistemic right: the ability to interrogate the categories, data proxies, and value judgments that shape algorithmic outputs [18]. Such a right operates as a corrective to testimonial injustice, reinstating the patient’s credibility when it is displaced by the apparent neutrality of machine authority. It also responds to hermeneutical injustice by creating space for alternative interpretive resources to enter the clinical dialogue. A patient who feels misrecognized by an algorithm should not be compelled into passive acquiescence but should be empowered to articulate why the system’s framing fails to capture their lived experience. For this to be meaningful, institutions must cultivate environments where contestation is not penalized or dismissed as irrational resistance but is treated as a legitimate epistemic contribution to the care process.

Alongside contestation stands the right to explanation. Existing debates on explainable AI have often been framed in terms of technical transparency, focusing on disclosure of model logic, feature weights, or statistical correlations [2]. Yet such technical disclosures are of limited value to patients and often insufficient even for clinicians without specialized training. What is required instead is what we might call a relationally meaningful right to explanation [44]. Explanations should be accessible, contextual, and actionable, enabling co-interpretation rather than mere compliance. For patients, this means that explanations must be connected to their embodied experiences and well-being, showing how an algorithmic recommendation relates to what they know and feel about their own condition. For clinicians, explanations must support professional judgment, allowing them to integrate algorithmic insights with patient testimony and clinical expertise rather than replace them. In this way, explanation becomes not a technical supplement but a relational practice, reinforcing trust and dialogue rather than eroding them. This relational stake is not merely theoretical: recent qualitative work comparing LLM-generated ethical guidance across cases finds that models diverge markedly in cultural sensitivity and interpretive depth despite invoking the same formal principles, underscoring that whose interpretive frame gets encoded remains consequential even when systems are framed as collaborative rather than authoritative [33].

The second dimension of the framework concerns the structural embedding of epistemic justice. Epistemic injustice in healthcare does not occur solely in the dyadic encounter between patient and clinician; it is shaped upstream by how categories are constructed, which narratives are included in training data, and how systems are evaluated. If algorithmic gaslighting arises from erasures in data, then it is not enough to refine model accuracy; the epistemic structures themselves must be rethought. One way to approach this is through adversarial auditing, understood here not as a compliance mechanism but as a normative practice of critical reflection. To examine training data and outputs for representational exclusions, to probe algorithms for systematic misrecognition of marginalized groups, is to enact a form of epistemic responsibility that resists the normalization of erasure [56]. Such audits should not be confined to technical experts but should include ethicists, patient advocates, and those with lived experience of marginalization. Only in this way can they surface not just statistical imbalances but the deeper hermeneutical gaps that algorithms encode.

A further structural commitment lies in participatory design (PD). Hermeneutical injustice arises when the interpretive resources available to a community fail to capture its experiences, leaving individuals unable to make sense of or communicate their suffering [24]. In healthcare AI, this injustice is amplified when systems are designed without input from those most likely to be misrecognized. PD methodologies address this by involving affected users in shaping the categories, narratives, and prototypes from the earliest design stages, rather than consulting them only after a system’s architecture is fixed [17]. This is not merely a pragmatic step to improve usability but a normative commitment to epistemic pluralism: the recognition that diverse forms of knowledge, including embodied and experiential knowledge, must have a place in the constitution of medical systems.

That said, PD in healthcare is not a uniformly available remedy, and treating it as one risks the same erasure the article otherwise criticizes: recommending a method that only well-resourced, low-friction settings can actually implement. A systematic review of PD in digital health development finds that even where PD is used, the scope, intensity, and stage of stakeholder involvement vary enormously across projects, and that sustained, iterative participation — as opposed to a single consultative session — remains the exception rather than the rule [71]. Several structural features of healthcare make sustained PD difficult to realize. Proprietary intellectual property protections over model architecture and training data limit vendors’ willingness to expose design choices to outside stakeholders at all, let alone iteratively. Privatized and fragmented health systems create no single institutional actor with both the authority and the incentive to fund long-term participatory processes, particularly where the commercial developer, the purchasing hospital system, and the treating clinician are three separate organizations with three separate interests. Acute and emergency care settings, where algorithmic decision support is often most consequential, are also the settings least amenable to participatory iteration, since the time pressures of the clinical encounter leave little room for the kind of open-ended co-design workshops PD typically requires.

PD is accordingly best understood as more feasible in some contexts than others, and the article’s recommendation should be read accordingly. It is most tractable in publicly funded or single-payer systems, in chronic disease management programs where patients interact with a system repeatedly over time, and in conditions with organized patient-advocacy infrastructure already in place — long COVID, chronic pain, and disability communities being examples where patient organizations already exist as ready interlocutors for design processes. It is least tractable for proprietary, vendor-controlled diagnostic tools sold across many purchasing institutions, and for acute-care algorithms operating under severe time constraints. Where full PD is infeasible, lighter-weight mechanisms — structured advisory input during procurement, or the “epistemic justice checks” described below — may be more realistic partial substitutes, though they do not substitute for genuine co-design where it is achievable.

The question of representation is equally pressing and equally under-specified in prior PD literature: which patients should stand in for a given condition or community? No single patient can represent the full range of experience within a diagnostic category, and treating one participant as representative of, say, “women with chronic pain” or “Black maternal health patients” risks reproducing exactly the kind of interpretive flattening this article critiques. A defensible approach requires, at minimum, intersectional sampling that avoids collapsing a heterogeneous patient population into a single voice; rotating or ongoing participation rather than one-time consultation, so that design commitments can be revisited as the technology and its user base evolve; and structures of compensation and shared authority — rather than uncompensated advisory labor — so that participation is not itself an extractive relationship of the kind this article otherwise criticizes algorithmic systems for enacting. When marginalized voices shape the hermeneutical resources of the algorithm itself under these conditions, the risk of gaslighting is mitigated at its source; where these conditions cannot be met, that limitation should be acknowledged rather than assumed away.

Structural commitments also extend into institutional adoption and governance. If algorithmic gaslighting operates by recalibrating epistemic agency, then procurement and deployment processes must explicitly include epistemic justice as a criterion. One way to conceptualize this is through what might be termed “epistemic justice checks” [36, 73]. Such checks are not simply compliance tick-boxes but moments of normative reflection: do patient-facing interfaces solicit narrative input rather than reducing testimony to checkbox data? Are explanation features designed to foster shared understanding rather than technical transparency alone? Has system performance been validated across diverse contexts, ensuring that marginalized experiences are not erased? By embedding such questions into institutional routines, healthcare organizations can treat epistemic justice not as an afterthought but as a constitutive element of adoption.

The third dimension of the framework concerns responsibility. A persistent challenge in AI ethics has been the diffusion of accountability across multiple actors, often described as the “accountability gap” [60]. In healthcare, this gap is particularly acute: developers build the models, institutions procure them, clinicians use them, and regulators oversee them, yet no single actor appears responsible when epistemic harms occur. Addressing algorithmic gaslighting requires moving beyond this fragmentation toward a shared responsibility model. As Section “Organizational intentionality and the politics of standardization” argues, institutions themselves bear a distinct form of responsibility prior to deployment, since the choice to standardize clinical information in case-based form is an intentional organizational act rather than a neutral technical default. Developers bear responsibility for documenting model limitations and designing for contestability. Clinicians bear responsibility for integrating patient testimony with algorithmic outputs rather than deferring uncritically to the latter. Institutions bear responsibility for creating governance structures—such as epistemic justice committees—that monitor deployment and mediate disputes. Regulators bear responsibility for ensuring that epistemic justice is codified as a standard of care. None of these responsibilities can be discharged in isolation; they must be coordinated within a framework of shared moral commitment.

At its deepest level, the framework calls for a moral commitment to epistemic justice as a constitutive value of healthcare. This requires moving beyond compliance-driven ethics toward a recognition that patients are not merely objects of knowledge but epistemic agents whose self-trust, credibility, and interpretive autonomy must be preserved. Such a commitment resonates with broader traditions in care ethics, which emphasize relationality, mutual respect, and the co-construction of meaning [69]. To design and deploy healthcare AI without this commitment is to risk reproducing the very structures of dismissal, distortion, and erasure that feminist and disability bioethics have long critiqued in human-only medicine. To design with this commitment is to create systems that expand, rather than contract, the epistemic capacities of patients, thereby reinforcing the possibility of trustworthy and humane medicine.

The proposed normative framework for addressing algorithmic gaslighting in healthcare operates on three levels: restoring individual agency through rights to contestation and meaningful explanation; embedding structural practices such as critical audits, participatory design, and epistemic justice checks; and cultivating a model of shared responsibility that treats epistemic justice as a foundational value. What unites these levels is a refusal to treat algorithmic gaslighting as a technical aberration reducible to bias or opacity alone. Instead, algorithmic gaslighting is understood as a structural and relational configuration in which epistemic injustice becomes infrastructurally mediated and experientially internalized, demanding structural and relational remedies. By centering epistemic justice, this framework shifts the focus from performance optimization to the preservation of patients as knowers, ensuring that the integration of AI into healthcare does not erode but strengthens the conditions of trust, dialogue, and agency.

Conclusion

The articulation of algorithmic gaslighting in healthcare underscores that algorithmic systems are not neutral tools but active participants in shaping credibility, interpretation, and trust in clinical practice. Drawing on feminist analyses of medical gaslighting and situated within digitally mediated care, this concept highlights how established forms of epistemic injustice can become stabilized within algorithmic infrastructures.

The analysis has shown that these harms are not accidental. Section “Conceptual foundations” situated them within the framework of epistemic injustice; Section “Mechanisms and harms of algorithmic gaslighting in healthcare” examined the sociotechnical mechanisms—data violence, opacity, automated authority, and feedback loops—through which they are operationalized; and Section “A normative framework for interrupting algorithmic gaslighting in healthcare” developed a normative response centered on epistemic justice. Taken together, these sections clarify how testimonial and hermeneutical injustices, when mediated by algorithmic systems, can assume a structurally reinforced and experientially internalized form within healthcare. The concern is therefore not limited to technical error, bias, or transparency deficits alone, but to the broader relational conditions under which epistemic agency is negotiated.

This analysis has been intentionally conceptual and normative in scope. It does not provide an empirical survey of deployed systems nor quantify the prevalence of the harms described. Rather, its contribution lies in offering a conceptual vocabulary for examining how epistemic injustice operates within digitally mediated clinical environments. Empirical research will be essential for tracing how these dynamics unfold across specific domains, but such inquiry benefits from a clarified normative framework capable of identifying what is ethically at stake.

If left unaddressed, these patterns risk entrenching mistrust, marginalization, and epistemic objectification in medicine. If, however, epistemic justice is treated as a constitutive value—through contestation rights, participatory design, critical auditing, and shared responsibility—algorithmic systems can be oriented toward strengthening rather than undermining patient agency. The central ethical question is therefore not whether healthcare will be digitally mediated, but how such mediation can preserve and support the epistemic standing of those whose lives it seeks to improve.

Author contributions

The sole author was responsible for the conceptualization, analysis, writing, and final approval of the manuscript.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This article presents a conceptual philosophical analysis and does not involve human participants, animal subjects, plants, or primary data collection. No ethical approval was required. The author declares no competing interests.

Competing interests

The authors declare no competing interests.