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.

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.

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].

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].

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%).

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.

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 3 Multimedia Appendix 3). Medical records and health information technicians (65%), nurse midwives (50%), and pediatricians (43.75%) had the highest task adoption rate.

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.

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.

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.

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.

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.

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.