Evaluating User Interactions and Adoption Patterns of Generative AI in Health Care Occupations Using Claude: Cross-Sectional Study

Introduction

Generative artificial intelligence (GenAI) systems, developed by organizations such as Anthropic (Claude), OpenAI (ChatGPT), Google (Gemini), Microsoft (Copilot), and Meta (Llama), represent a significant technological shift and are rapidly being adopted in diverse sectors globally at an unprecedented speed [1]. Health care represents a domain where GenAI holds considerable transformative potential. Proposed applications range from enhancing clinical decision-making support [2,3] and potentially automating high-skilled cognitive tasks distinct from previous software capabilities [4], to streamlining burdensome administrative workflows and empowering patients with more accessible, tailored health information [2,3]. The rapid proliferation and increasing sophistication of these tools necessitate a clear understanding of their current role and impact within the health care ecosystem.

Initial research has started to explore the capabilities of GenAI in health care contexts, often demonstrating performance on par with, or even exceeding that of human clinicians on specific, well-defined tasks such as diagnostic reasoning or medical knowledge recall [5-7]. For instance, studies have compared GenAI outputs to physician performance in complex diagnostic challenges [6] and clinical reasoning assessments [5]. Furthermore, user trust plays a critical role in the adoption of these AI tools [8], adding another layer to understanding real-world uptake. However, these studies typically occur in controlled or experimental settings, focusing on benchmark performance rather than quantifying the extent and nature of GenAI’s spontaneous, real-world application across the breadth of health care activities.

Consequently, a significant gap exists between understanding GenAI’s potential capabilities and characterizing its actual adoption patterns and usage frequency for health care–related tasks by both professionals and the public. Studies have confirmed that rapid AI adoption is occurring across the economy [1,9], highlighting the urgency of understanding its domain-specific implications. This widespread availability and adoption of GenAI tools raise many questions about their real-world integration [10]. As individuals increasingly turn to free and widely available GenAI models, how frequently are these interactions related to health care? Are certain clinical roles or types of health care tasks experiencing disproportionately high engagement via GenAI? Given these questions, we sought to move beyond capability assessment and provide empirical data on the current landscape of GenAI usage for tasks associated with health care occupations.

Using Anthropic’s foundational “Economic Index” research which provided a broad overview of artificial intelligence (AI) usage across various economic sectors [11], this cross-sectional study aimed to conduct a more focused analysis specifically within the health care domain. Leveraging the same large-scale, anonymized dataset of user interactions with Claude, our primary objectives were to (1) quantify the overall proportion of GenAI interactions dedicated to tasks associated with health care occupations, as defined by the US Department of Labor’s [12] O*NET database [13]; (2) identify which specific health care roles exhibit the highest frequency of associated task interactions; (3) introduce and assess a novel metric—the “digital adoption rate,”—representing the breadth of GenAI adoption within these roles based on the variety of distinct occupational tasks performed using the AI; and (4) contextualize health care–related usage by comparing it with usage patterns across other major occupational sectors.

Methods

Data Source and Study Design

We leveraged a publicly available, privacy-preserving aggregated dataset derived from over 4 million anonymized conversations with Claude from free and pro subscribers, released as part of Anthropic’s “Economic Index” research [11]. This dataset reflects a broad spectrum of real-world user interactions, offering an ecologically valid basis for examining GenAI engagement with professional tasks.

Occupational Task Mapping

Our analysis used the premapped dataset created and open-sourced by Anthropic, which contains conversations already mapped to standardized occupational tasks as defined by the US Department of Labor’s O*NET database. Anthropic’s team used their proprietary Clio model to create these mappings and validated them through manual review [14]. Specifically, they used hand validation across 150 examples for task hierarchy classifications, finding that 86% of conversations were judged as correctly assigned at the base O*NET task level, 91.3% at the middle level, and 95.3% at the top level of their hierarchical framework [11]. We focused our analysis specifically on the subset of this dataset related to tasks and occupations within the health care domain. Figure 1 illustrates the Clio mapping methodology that underpins the dataset in this study.

‎

Metrics and Outcomes

• Occupation-specific GenAI usage: We computed the proportion of total GenAI conversations attributable to each O*NET Healthcare Practitioners and Technical Occupation.
• Global occupational categories: We compared health care’s share of GenAI usage against other occupational categories (eg, computing, arts and media, education) to contextualize overall representation.
• Digital adoption rate: We calculated the ratio of distinct tasks performed with GenAI to total distinct potential tasks for each occupation. For instance, if an occupation had 20 potential tasks in the O*NET database and 5 of them appeared in GenAI conversations, the adoption rate was 25%. This analysis was conducted for the top 25 health care practitioners and technical occupations.

Statistical Analysis

Descriptive statistics summarized GenAI usage proportions and adoption rates. We used simple comparisons (eg, ratio of usage in health care vs other categories) to maximize interpretability while highlighting relative differences. All analyses were performed using R software (version 4.3.1; R Foundation for Statistical Computing) [15].

Ethical Considerations

Ethics Committee of The Ohio State University determined that our study did not qualify as human subjects research requiring review as no identifiable personal data were included in the dataset. All conversations were preanonymized and aggregated, and the data were made public by Anthropic on February 10, 2025 [16].

Results

Overall Health care Usage

Tasks mapped to Health care Practitioners and Technical Occupations collectively represented 2.58% of all GenAI conversations (Table 1 and Multimedia Appendix 1). In contrast, computing‐related fields accounted for 37.22%, followed by arts and media (10.28%) and education (9.33%).

Within-Health Care Breakdown

In the top 25 list of Health care Practitioners and Technical Occupations by GenAI conversation share, dietitians and nutritionists (6.61%), nurse practitioners (5.63%), music therapists (4.54%), and clinical nurse specialists (4.53%) had the highest usage (Table 2 and Multimedia Appendix 2). Meanwhile, other occupations such as pediatricians (1.38%) and internists (1.29%) showed more modest representation.

Adoption Rates

Digital adoption rates for top health care occupations ranged from 13.33% to 65%, with an average of 16.92%, compared to a global task adoption rate of 21.13% (Table 3Multimedia Appendix 3). Medical records and health information technicians (65%), nurse midwives (50%), and pediatricians (43.75%) had the highest task adoption rate.

Discussion

Interpreting GenAI Usage Patterns in Health care Tasks Through Multiple Perspectives

Our analysis reveals significant GenAI interaction volume related to health care tasks, yet concentrated within specific roles, particularly those involving substantial patient interaction, education, and guidance (eg, dietitians, nurse practitioners, therapists). These findings point toward an emerging, complex integration of GenAI into the health care ecosystem. However, the anonymized nature of the data necessitates consideration of multiple interpretations about the users driving this usage and their underlying reasons.

Interpretation 1: Patient-Driven Information Seeking and Self-Management

One plausible interpretation is that the observed patterns reflect significant usage by the public. Patients may be turning to GenAI such as Claude, for accessible health information, clarification of medical terminology, lifestyle advice, or even emotional support related to health concerns—tasks commonly associated with roles such as dietitians or nurse practitioners. Such information-seeking behavior aligns with pre-existing trends of patients consulting “Dr. Google” [17] and reflects a potential desire for readily available information [18]. Observed patterns among ChatGPT users seeking health information, who tend to be younger, report poorer health status, and more frequently use transient care settings [19], may further point to this behavior, particularly when timely access to health care professionals is limited [20]. If patients are indeed primary users, this trend could signify a shift toward greater patient engagement and potential empowerment through readily accessible, personalized information. However, this interpretation also carries significant risks: the potential for consuming misinformation or AI “hallucinations,” leading to misguided self-management decisions, delayed consultation with qualified professionals, or increased health anxiety [18]. These risks have been highlighted in studies examining user trust and reliance on chatbots for health advice [8]. The accessibility that drives usage could simultaneously introduce harm if not accompanied by subject matter expertise, digital health literacy, and critical appraisal skills.

Interpretation 2: Professional Workflow Assistance and Efficiency

Alternatively, the concentration of usage in these roles could indicate adoption by health care professionals themselves. Dietitians, nurse practitioners, clinical nurse specialists, and others might be leveraging GenAI to enhance their workflow–drafting patient education materials, summarizing complex information, exploring differential diagnoses, generating notes, or staying abreast of new research relevant to patient counseling. These uses suggest potential automation of specific cognitive tasks previously resistant to software solutions [4]. From this perspective, GenAI acts as an efficiency tool, potentially helping overloaded professionals manage demanding workloads within a strained health care system. Indeed, workers in exposed occupations often perceive substantial productivity potential in GenAI tools like ChatGPT [9]. This interpretation suggests a pathway towards augmenting professional capabilities, potentially improving the consistency or breadth of patient communication. Yet, risks exist here as well, including over-reliance on AI outputs that potentially contain inaccuracies or biases, deskilling in core communication or critical thinking tasks through cognitive offloading, and ethical implications of using AI to generate content in patient care without adequate oversight or personalization.

The Underlying Influence of Health care System Pressures

Regardless of whether patients or professionals are the primary users, the observed patterns may also indicate broader systemic issues within health care. When professional support is not easily or quickly accessible, individuals—both patients and potentially time-pressed professionals—are likely to turn to readily available tools such as GenAI. Findings that users seeking health information from ChatGPT tend to report poorer health and greater use of transient care settings lend credence to this possibility [19]. This suggests that GenAI adoption in health care contexts might be partially driven by unmet needs and access gaps within the current system. Furthermore, even when productivity benefits are perceived, adoption may be hindered by factors such as employer restrictions or the need for training [9]. This “compensatory” use stresses the importance for health care systems to adapt, not only by integrating AI safely, but by addressing the root causes that drive patients or professionals toward potentially unverified AI tools out of necessity.

Implications for the Patient-Provider Dynamic and Information Hierarchy

The convergence of these possibilities, patients potentially arriving at appointments armed with AI-generated information, and professionals using GenAI to prepare for those encounters, points toward a fundamental reshaping of information hierarchies and the patient-provider relationship. The dynamic shifts from a traditionally asymmetric flow of information to one where both parties may engage with AI-derived knowledge. This shift demands new communication strategies, enhanced critical appraisal skills from both sides, and a shared understanding of GenAI’s current capabilities and limitations [7]. This is particularly relevant given that a significant portion of patients report acting on AI-generated health information, such as changing medications or seeking referrals, and often subsequently discussing the AI output with their physicians [19]. Addressing the potential for GenAI to exacerbate existing health inequities, especially those based on digital access or literacy, is also paramount.

Future Directions

Health care organizations and regulatory bodies should develop comprehensive frameworks for responsible AI integration. These should include creating evidence-based guidelines for clinical practice integration and establishing rigorous standards for the accuracy of AI-generated health information. Ethical considerations surrounding data privacy, training bias, and the potential for exacerbating health inequities must also be central to any framework for responsible GenAI integration. Future research should investigate whether using GenAI translates into improved health outcomes and how shifts in digital literacy affect patient engagement. Importantly, developing methodologies to better differentiate between professional and lay user interactions with health-related GenAI queries will be essential for accurately interpreting adoption trends and tailoring interventions at scale.

Limitations

First, we relied on anonymized, aggregated conversations; thus information on user credentials and demographics is not available. Second, the mapping from conversation text to occupational tasks may be imperfect, particularly when laypersons seek professional-grade information. Third, our cross-sectional design provides a single snapshot in time rather than longitudinal trends. Fourth, data from conversations were limited to Claude since Anthropic is the only company that has open-sourced data for analysis. The demographics of Claude users, along with the reasons why users engage with Claude, may or may not be representative across other GenAI tools which could fundamentally bias usage patterns. Finally, we do not know what users did with the information they received from GenAI or how it impacted health outcomes. Future analyses should account for ongoing changes in GenAI capabilities and usage patterns.

Conclusion

Generative AI is making tangible inroads into health care–related tasks, yet its adoption appears selective, with high interaction volumes associated with patient-facing roles such as dietitians and nurse practitioners. Our analysis, based on interactions with Claude, highlights this emerging pattern but calls attention to a critical ambiguity: whether this usage is primarily driven by patients seeking information and navigating access challenges, by professionals seeking workflow efficiencies, or by a combination of both.

Regardless of the primary user, this integration demands immediate attention. The observed patterns necessitate a shift beyond simply acknowledging GenAI’s potential, toward implementing proactive strategies. Future research must focus on identifying who uses these tools for health care–related tasks and understanding their motivations, potentially through methods that more accurately infer user intent or role. Concurrently, stakeholders including clinicians, educators, developers, and policymakers must develop distinct guidance—resources promoting digital health literacy and critical appraisal for patients, and evidence-based best practices for safe and effective integration into professional clinical workflows. Addressing how systemic health care pressures may be influencing this adoption is also crucial. Ensuring GenAI enhances, rather than compromises, health care delivery requires targeted action now to shape its responsible use before usage patterns solidify.