Rare diseases collectively affect an estimated 10% of the global population, corresponding to approximately 500 million people worldwide [1]. Although individually uncommon, rare diseases constitute a large and heterogeneous group, with approximately 7000 distinct conditions currently described [2]. However, the distribution of cases is highly skewed: roughly 350 rare diseases account for around 80% of diagnoses, while thousands of additional conditions are classified as ultrarare [2]. Definitions of rarity vary internationally, with different taxonomic thresholds, including fewer than 1 in 2000 individuals in the European Union, fewer than 200,000 people in the United States, and fewer than 50,000 people in Japan [1,3]. In addition, genetic rare diseases differ across geographic regions, making the challenge of clinician diagnoses even greater in many cases. Together, these figures illustrate that rare diseases are neither exceptional nor marginal in aggregate, yet they remain structurally underrecognized within health care systems.

People with rare diseases frequently experience prolonged and complex diagnostic pathways. Substantial delays between symptom onset and correct diagnosis are common. A study in Australia found that approximately 30% of individuals waited between 5 and 30 years for a definitive diagnosis [4]. In the United States, another study reported that around 50% of patients and caregivers attribute diagnostic delays to limited clinician awareness of their condition [5]. These challenges occur within a clinical environment characterized by the rapid expansion of biomedical knowledge. In addition to thousands of existing rare diseases, approximately 250 new rare diseases are described each year, accompanied by the continual refinement of gene-disease associations and molecular subtypes [2]. Under these conditions, it has been argued that maintaining up-to-date familiarity across rare disease domains is not realistically achievable for individual clinicians [6].

Even when an accurate diagnosis is achieved, therapeutic options are often limited. Fewer than 10% of rare diseases currently have an available disease-modifying treatment, and approximately 95% lack a US Food and Drug Administration–approved therapy [2]. Evidence generation is constrained by small patient populations, fragmented research efforts, and high rates of trial noncompletion or nonpublication [7,8]. As a result, patients and families commonly engage in sustained self-directed information seeking, including monitoring emerging research, identifying potential clinical trials, and synthesizing highly technical biomedical information to support care decisions [9].

In parallel, rare disease care has long challenged conventional assumptions about where medical expertise resides. In part, because of knowledge deficits among clinicians and diagnostic delays, patient-researchers argue that people living with chronic and rare conditions often develop deep, experience-based expertise and may actively contribute to research, innovation, and care improvement rather than merely receiving it [9,10]. However, health care systems frequently struggle to recognize patient and caregiver knowledge. Philosophers describe this problem as one of “epistemic injustice”—a “wrong done to someone in their capacity as a knower,” in which patients’ accounts tend to be granted less credibility than professional perspectives [11,12]. In clinical settings, this can manifest as “testimonial injustice” (patients not believed) or “hermeneutical injustice” (patients’ experiences not fitting existing medical categories), phenomena closely related to what patients often describe as medical gaslighting [11,13]. Set against these concerns, the rapid public availability of large language model (LLM)–powered chatbots has burgeoned when it comes to health-related searches [14,15]. The use of consumer generative artificial intelligence (GenAI) tools, such as ChatGPT, Gemini, and DeepSeek, and a range of other models has created new possibilities for processing health information. Public access to consumer GenAI tools is a recent phenomenon. ChatGPT, the first widely available LLM chatbot, was released by OpenAI in November 2022. This prompted the rapid release of competing tools, including Google’s Gemini (2023) and Anthropic’s Claude (2023), among others. GenAI tools are, therefore, readily available but are susceptible to producing inaccurate or fabricated outputs and may encode or amplify existing biases, including those related to race, gender, and disability, with potential consequences for equity in health care [16]. In addition, the widespread availability of consumer-facing GenAI tools, coupled with their highly interactive conversational interfaces, raises concerns about privacy, data security, and the handling of sensitive personal health information [17,18].

To date, empirical research on GenAI in health care has primarily examined clinician adoption of both consumer and medical-grade applications for a range of clinical tasks, including documentation support, clinical reasoning, and treatment recommendations [19]. In contrast, far less is known about how patients and family members use GenAI tools [19-22]. This gap is particularly consequential for rare disease communities, where informational needs are persistent and clinical expertise is unevenly distributed [20]. However, one US survey found that households with rare diseases were twice as likely to use GenAI tools compared with other households [23].

Moreover, even before GenAI, people affected by rare diseases were established users of online resources—both to manage the chronic uncertainty of conditions with life-limiting potential, such as cancer predisposition syndromes and degenerative disorders [24-26], and to connect with peers and advocacy organizations for social support [27-30]. Consumer GenAI tools thus enter an information ecosystem in which these communities are already highly engaged online.

Notably, parents in rare disease communities have a documented history of being early adopters of novel and at times controversial technologies [31], suggesting that engagement with emerging GenAI tools is consistent with longstanding patterns of technology uptake in this population. This study takes epistemic injustice as its theoretical framework, asking how patient and caregiver GenAI use intersects with the credibility deficits and interpretive gaps common among patients with rare diseases, many of whom wait decades to be diagnosed [4,5].

Against the theoretical framework of epistemic injustice and the research gaps related to patient use of GenAI, the objective of this mixed methods study was to examine the experiences and opinions of adult patients with rare diseases and parents or guardians of children with rare diseases regarding the use of GenAI tools. Specifically, we asked the following questions: How do patients and family members in rare disease communities use GenAI tools? What purposes do these tools serve? What benefits and concerns are perceived? To what extent do participants report that AI-generated information influences health care decisions or interactions with clinicians? We addressed these questions using a convenience sample recruited through rare disease networks and community organizations in the United States, combining quantitative survey data with open-text qualitative responses.

This study used an exploratory mixed methods cross-sectional survey design to examine the experiences and opinions of patients with rare diseases and parents or guardians of children with rare diseases regarding the use of GenAI tools in the context of rare disease care. A mixed methods approach was selected to enable descriptive quantification of patterns of use alongside qualitative exploration of concerns and contextual experiences. Given the limited existing patient-centered evidence in this area, an exploratory design was considered appropriate.

The study received ethics approval and was deemed exempt from the human participants research requirements of 45 CFR 46 by the Boston Children’s Hospital Institutional Review Board (approval ID IRB-P00051505). Electronic informed consent was obtained from all participants prior to survey initiation. Participation was voluntary, and respondents could withdraw at any time before submitting the survey. In addition, no compensation was provided for completing the survey.

An original survey instrument was developed by the study team to address the study objectives. Survey content was informed by prior literature on patient information-seeking behaviors, emerging scholarship on GenAI in health care, and consultation with rare disease community stakeholders [23,32,33]. We distinguish here between our patient partner (CEB), who contributed to survey co-design, and the survey respondents (N=115), who participated anonymously. Specifically, our patient partner (CEB) with lived experience of a rare disease contributed to item development, wording, and assessment of relevance. Draft versions of the survey were reviewed by contacts within participating rare disease networks to assess clarity, accessibility, and face validity, and the instrument was timed to ensure feasibility of completion within approximately 5 minutes.

The final survey comprised 19 primary items, including multiple-choice, Likert-type, and checkbox questions, together with 4 standalone open-text free-response questions (items 10a, 16a, 17a, and 18). Open-text questions did not accompany every closed-ended item; rather, item 10a invited respondents to specify other purposes of GenAI use, whereas items 16a, 17a, and 18 invited extended reflection on experiences of harm, factors influencing cautious or nonuse, and final reflections on rare disease care and the potential role of GenAI tools, respectively. The 3 additional reflection questions were as follows:

All questions were optional. Because the study did not aim to examine identity-based differences and sought to reduce the collection of potentially sensitive personal data, gender was not included among the demographic items. The full survey instrument and informed consent materials are provided in Multimedia Appendix 1. The study adhered to the CHERRIES (Checklist for Reporting Results of Internet E-Surveys) guidelines [34] (Checklist 1).

Participants were eligible if they were aged 18 years or older and self-identified as either a patient with a rare disease or a parent or guardian of a child with a rare disease. No additional exclusion criteria were applied. Recruitment used convenience sampling through the established rare disease networks and community organizations COMBINEDBrain [35] and EveryLife Foundation [36]. Based in the United States, both are umbrella networks of rare genetic disease communities with validated patient and family member contacts. Members of these organizations were invited by email to distribute an anonymous survey link through their email lists and closed-network newsletters, and with 2 reminder emails requested during the recruitment period. To intentionally reduce the risks of bot contamination, duplicate participation, and unverifiable eligibility, the survey team and contacts were requested not to share the survey link on any open social media platforms.

The survey was hosted on REDCap (Vanderbilt University), which is a secure web-based data capture platform. Data collection occurred between November 1, 2025, and January 10, 2026. No personally identifiable information, including names, email addresses, or IP addresses, was collected, and responses were anonymous. In addition to fixed-response items, as noted previously, participants could provide open-text responses describing other purposes for GenAI use, experiences of harm or negative outcomes, additional factors influencing cautious use or nonuse, and any further reflections on rare disease care or the potential role of GenAI tools. To reduce the risk of reidentification, particularly given the rarity of some conditions and the potential for triangulation with demographic information, we did not share row-level raw data. Instead, we provided the full aggregated and tabulated responses in Multimedia Appendix 2.

Quantitative data were analyzed using descriptive statistics, including frequencies and percentages, to summarize respondent characteristics, prior use of GenAI, purposes of use, perceived influence on medical decisions, trustworthiness, concerns, and communication with health care professionals. Given the exploratory nature of the study and the limited expected sample size, no inferential statistical analyses were planned.

Qualitative thematic analysis was restricted to the 3 reflective free-response items (items 16a, 17a, and 18). Responses to item 10a (other purposes of use) consisted of brief specifications that supplemented the quantitative data on purposes of use and were not subjected to thematic analysis. Open-text responses were analyzed using a qualitative descriptive approach with inductive thematic coding, allowing patterns and concepts to be identified directly from the data rather than from pre-existing frameworks [37,38]. However, given the concise nature of many responses and the frequent use of brief phrases or sentence fragments, the data were not suited for in-depth interpretive thematic analysis [39]. Prior to analysis, responses indicating no substantive content (eg, “none,” “not applicable,” or “no comment”) were removed. The remaining responses were imported into an Excel file for data management and coding.

Two members of the research team (CB and CEB) independently reviewed the responses multiple times to develop familiarity with the dataset. CB is a philosopher, ethicist, and informaticist, and CEB is her twin sister, a government worker, who lives with a rare illness. An inductive coding strategy was then applied, whereby short descriptive codes were assigned to capture the central idea expressed in each response. Responses that conveyed more than one idea were assigned multiple codes. Coding was iterative, with ongoing comparison of codes to identify areas of convergence and divergence. CB and CEB met to discuss coding decisions, resolve only minor and trivial discrepancies, and refine the coding framework. Given the short and often fragmentary nature of the open-text responses, formal intercoder reliability statistics and thematic saturation were not appropriate because these techniques presuppose richer narrative data than this instrument was designed to elicit. Final codes were subsequently grouped into higher-order categories based on conceptual similarity, generating a structured summary of participant perspectives. Again, this approach is entirely consistent with the limitations of the dataset [39] and has been undertaken in previous analyses by the study team [40]. Owing to the online survey format and the limited responses, we were unable to establish thematic saturation.

A total of 115 respondents completed the survey. The majority of respondents (n=74, 64.3%) identified as parents or guardians of a child with a rare disease, whereas the remainder (n=41, 35.7%) identified as patients with a rare disease (Table 1). Participants were predominantly adults, with 61.7% (n=71) aged between 35 and 54 years. Educational attainment was high, with 75.7% (n=87) reporting a university or postgraduate degree. Most respondents resided in the United States (n=98, 85.2%). A breakdown of the results by respondent group is further presented in Multimedia Appendix 3.

Respondents reported consulting a mean of 7.3 (SD 12.0) health care professionals prior to receiving a correct diagnosis (Table 2). Diagnostic timelines were frequently prolonged, with 43 of 115 (37.4%) reporting a wait of 3 years or longer, including 16 of 115 (13.9%) who reported delays exceeding 10 years. Among the participants, 22 of 115 (19.1%) reported receiving an official rare disease diagnosis since 2023.

Slightly more than half of the respondents (63/115, 54.8%) reported some prior use of GenAI tools for rare disease–related purposes, whereas 52 of 115 (45.2%) respondents reported never using such tools (Table 3). Among users, the most commonly reported purposes were exploring new treatments or clinical trials (53/63, 84.1%) and interpreting medical tests or clinical notes (37/63, 58.7%). One-third of respondents (37/115, 32.2%) reported additional uses, including summarizing scientific publications and supporting insurance appeals or advocacy.

As shown in Table 4, most respondents reported minimal or no influence of GenAI tools on medical decisions (77/115, 67.0%). However, nearly one-third of respondents (37/115, 32.2%) reported that GenAI had influenced their medical decisions to some degree, including 10 (8.7%) respondents who reported a strong influence. Direct contributions of GenAI to formal diagnoses were uncommon (11/115, 9.6% reporting any contribution), and most respondents reported that GenAI had not shortened their diagnostic journey (93/115, 80.9%).

Perceived trustworthiness of AI-generated health information was mixed: of 115 respondents, 45 (39.1%) respondents rated it as less trustworthy than traditional sources, while 33 (28.7%) rated it as equally trustworthy (Table 5). Concern about accuracy was widespread, with 71 (61.7%) respondents reporting moderate, very high, or extreme concern. Notably, most respondents (n=90, 78.3%) had not discussed AI-generated information with a health care professional.

Few respondents reported experiencing harm or negative outcomes related to GenAI use (7/115, 6.1%; Table 6). When considering factors that dissuade respondents from using GenAI tools, most cited uncertainty about accuracy (75/115, 65.2%), followed by privacy concerns (36/115, 31.3%) and lack of trust (35/115, 30.4%).

Currently, there is very little research on patient use of GenAI tools, particularly among people affected by rare diseases, a historically marginalized population that must often navigate care amid limited clinician expertise. In this convenience sample, mixed methods survey (N=115), just over half of the respondents (n=63, 54.8%) reported using GenAI for rare disease-related tasks, most commonly exploring treatments or trials (n=53, 46.1%), interpreting medical tests or notes (n=37, 32.2%), identifying specialists (n=29, 25.2%), and considering possible diagnoses (n=28, 24.3%). Qualitative responses indicated these activities functioned as extensions of caregiving and medical self-advocacy—helping participants understand complex genetic information and prepare questions—rather than as attempts to replace clinicians.

Participants used GenAI cautiously and strategically. Most reported minimal or no influence on medical decisions (n=77, 67.0%), formal diagnostic contributions were uncommon (n=11, 9.6%), and concern about accuracy was widespread (n=71, 61.8%). Few had discussed AI outputs with clinicians (n=90, 78.3%), and harms were rarely reported (n=7, 6.1%). Against a backdrop of prolonged diagnostic effort (clinicians consulted: mean 7.3, SD 12.0; n=43, 37.4% waiting ≥3 y), GenAI primarily served as a cognitive support for navigating fragmented care and potentially constrained access to expertise—enabling preparation, verification, and self-advocacy in medicine while maintaining reliance on human clinical oversight.

A large probability-based US survey conducted earlier in 2025 found that people in households with members having a rare disease were disproportionately early adopters of digital tools, with 38% reporting the use of AI tools to learn about a condition or treatment options [23]. However, that study characterized adoption at the population level and did not examine how patients use AI, how they judge its reliability, or how it may shape clinical decision-making. This study complements and extends those findings by showing that engagement is not merely uptake but invokes reflective and structured practice: participants described verification strategies, bounded trust, and selective incorporation of GenAI into advocacy and preparation rather than replacement of clinicians.

Conversely, compared with our convenience sample, the probability sampling of the earlier national survey provides stronger population representativeness. However, by integrating quantitative patterns with qualitative accounts, this study demonstrates that AI use in rare disease contexts may function subtly in patient adoption, rather than merely as simple technological enthusiasm. Together, these studies suggest that rare disease communities may be early adopters and careful interpreters of GenAI.

This study contrasts with the narratives presented in other highly cited studies. For example, recent experimental work using standardized vignettes suggests that access to LLM tools does not automatically improve lay diagnostic or triage accuracy [43]. However, these findings should be treated with caution and should not be interpreted as evidence that patients “believe misinformation.” Rather, such findings may reflect limitations of one-shot vignette paradigms and the difficulty of translating conversational suggestions into a single committed decision, especially when users are not living with symptoms over time. Our study survey adds important nuance: participants reported more cautious and critical GenAI use than is often assumed in clinical and academic commentary. These results underscore the importance of avoiding deficit framings that inadvertently discount patient expertise [42,44].

The study findings should also be read in light of the sample composition. Most respondents were already diagnosed, and many—particularly parents and guardians—were reflecting retrospectively or on a child’s diagnostic pathway, often during a period when GenAI tools were not yet available. The finding that GenAI was not perceived to have shortened the diagnostic journey may, therefore, reflect the timing and structure of the participants’ experiences rather than a limitation of GenAI’s potential to do so—an open question that prospective studies of undiagnosed patients are better placed to address. Future studies designed to capture time since diagnosis directly, and powered for subgroup analysis, could valuably examine whether GenAI use is prompted by diagnoses. Similarly, future work should be powered to explore patient or caregiver age, gender, and education level to explore demographic differences in the adoption of GenAI tools.

The study findings also align with the literature on epistemic injustice in health care [11,12]. Participants used GenAI not to replace clinicians but to translate genetic reports, prepare arguments, and support advocacy during prolonged diagnostic pathways. In this sense, GenAI tools appeared to function as a form of informational support: resources that helped patients to articulate and validate observations that might otherwise be dismissed. However, further qualitative research is needed to substantiate this claim in depth (eg, interviews).

Prior work has emphasized listening to patient concerns as a prerequisite for meaningful clinical recommendations and highlighted how “spetspatients” (expert patients) can reshape health care through knowledge production [9,10]. The study results suggest GenAI may increasingly mediate this process, potentially reducing but also revealing epistemic asymmetries in rare disease encounters. Yet clinicians in rare disease care also operate under genuine, often irreducible, uncertainty [6,45]. Read against this, patient GenAI use can be understood not only as a corrective to epistemic injustice but as a parallel response to a shared epistemic predicament—reframing it as a resource for collaborative management of uncertainty rather than a challenge to clinical authority [46].

It is important to note that the expert role rare disease patients and families take on is often less a chosen empowerment than a burden of necessity, undertaken in the absence of accessible clinical expertise or established evidence. Combined with the epistemic injustice they may encounter if their resulting knowledge is dismissed [12,42,47], rare disease patients and caregivers are placed in a structurally difficult position: they are required to acquire and interpret specialist information, yet may be denied standing as credible knowers when they do so [13,42]. Patient and caregiver use of GenAI should, therefore, be read against this dual pressure, as a tool taken up to manage a burden that the health care system has, in many cases, failed to relieve.

Our findings also align with an established literature documenting the expertise that patients with rare diseases and caregivers develop through both offline networks (such as advocacy organizations and patient registries) and online communities, where peers exchange clinical experiences, navigate uncertainty, and contribute to research agendas [27,48]. Our data extend this picture by showing that GenAI is now being layered onto this existing expertise infrastructure. The conflict this surfaces is not new but is sharpened by GenAI’s reach—patients and caregivers continue to perform substantial epistemic labor that is often unrecognized by health systems, even as the tools available to them become more powerful. This underscores the need for explicit recognition of patient and caregiver expertise in the design and governance of rare disease care, including GenAI’s role within it [20].

Of particular note, the participants in our exploratory survey were cautious about the present use of GenAI, with few reporting harms, yet many believed that these tools could have helped them during earlier stages of rare disease diagnosis. These findings are not contradictory but reflect timing. As noted, GenAI became publicly available only in late 2022, whereas most respondents (93/115, 80.9%) were diagnosed earlier and could not have used it during their diagnostic journey. The retrospective belief that GenAI might have helped therefore reflects unmet need rather than an evaluation of its actual contribution. Given the lengthy diagnostic pathways reported in this study, and for rare disease patients, these attitudes may reflect an adaptive response to unmet clinical need. The nuanced reflections on uptake also challenge dominant narratives that frame patient use of GenAI mainly as a safety concern.

This study has several limitations. First, we used a nonprobability convenience sample, which limits generalizability. Participants were recruited through web-based rare disease networks and community organizations, and it is likely that respondents who chose to participate were more digitally literate and more interested in digital tools or GenAI than nonrespondents. Responses were likely biased in favor of highly engaged patients and caregivers who had affiliated with these rare disease networks for information and peer support. Notably, network membership in this context reflects community engagement rather than an advocacy role, and the sample should not be read as composed of designated patient advocates. This is, however, the population in which patterns of early GenAI use are most visible and most relevant for informing emerging patient-facing guidance. Recent national data indicate that rare disease households are disproportionately early adopters of these tools [23], suggesting that our sample represents the leading edge of community uptake rather than an isolated subgroup.

Notwithstanding, and second, our sample was highly educated and based mostly in the United States, and the sample size was modest and not powered to examine associations between respondents’ demographic characteristics and outcomes. In addition, it was not possible to establish thematic saturation due to the limitations of the dataset. Third, to minimize respondent burden and limit the collection of potentially sensitive personal data, the survey included a restricted set of demographic variables and did not collect information on gender, precluding analysis of potential gender-based differences. Fourth, although the qualitative data provided important contextual insights, open-text responses were typically brief, limiting the depth of interpretation and precluding more nuanced qualitative analyses. For example, items such as the use of GenAI for emotional or social support could have been interpreted either as using chatbots for therapy bots, or as seeking out information about services to support patients and caregivers. Furthermore, more targeted interviews or focus groups could disambiguate differing interpretations.

These limitations should be considered alongside important strengths. To our knowledge, this study represents one of the first in-depth patient-centered and family-centered surveys to examine experiences and opinions regarding GenAI use in health care, which focuses specifically on rare disease communities. In addition, when it came to reported rare diseases, our sample was highly heterogeneous. Recruitment was conducted exclusively through validated patient and family networks, rather than open social media platforms, reducing risks of contamination, ineligible participation, and automated or duplicate responses. That said, we acknowledge that 87% of respondents did not state where they learned about the survey; this was an optional question, however, and it may have been perceived as redundant by our participants. The mixed methods design enabled the integration of quantitative patterns with qualitative perspectives, providing a more comprehensive picture of how GenAI tools are currently being used and perceived outside clinical settings.

Taken together, these findings provide early, exploratory evidence of a growing but largely unexamined layer of patient-facing GenAI use. More robust research is needed, including studies using larger and more representative samples, longitudinal designs, and in-depth qualitative methods, such as interviews or focus groups, to better understand patient experiences, potential benefits, risks, and implications for clinical practice and policy. Furthermore, translation of these early findings to wider populations will require complementary work, including plain-language patient guidance, peer-led informational resources distributed through rare disease networks, and clinician communication training, alongside larger and more representative survey samples.

This study reveals a largely invisible layer of patient-facing GenAI use occurring outside clinical oversight. Patients with rare diseases and their caregivers were not replacing clinicians but were using AI to interpret information, prepare questions, and support advocacy developed through prolonged diagnostic uncertainty.

Although the findings are limited by a modest, highly educated convenience sample and therefore not broadly generalizable, the study provides timely and original insight into how a historically marginalized patient population is engaging with emerging technologies. Rather than overreliance, participants demonstrated cautious and critical use while recognizing AI’s potential to reduce diagnostic delays and informational scarcity. GenAI, therefore, does not merely introduce new risks into rare disease care; it also exposes longstanding epistemic gaps within it and suggests that future governance and clinical practice should be organized around partnership with informed patients rather than re-enacting justifications for paternalistic protectionism.