Participants pointed out that the widespread use of AI requires energy and human resources. Furthermore, AI requires substantial technical resources, which are currently not available ubiquitously, as the technical infrastructure, for example, the availability of software and hardware, varies between health care facilities, making widespread use and networking difficult. Participants supported regional and global networking for physicians and health care facilities to build databases and use AI systems, for example, to share information or coordinate treatment, but perceived the technical implementation as challenging. The lack of standardized software and incompatibility between existing systems were also regarded as challenges to the successful use of AI and networking. They emphasized the importance of compatibility between newly developed systems and those already in use.

Therefore, participants considered balancing the provision of technology, and in particular, its practical functionality and maintenance, as a major challenge. Ensuring energy security and rapid technical service in the event of system failure was perceived as challenging but crucial, especially for medical applications on patients. Participants noted the need for human resources to develop and test AI, considering it difficult if additional staff were required to analyze and operate the AI, given the current staff shortage in the health sector. Participants also highlighted the importance of environmental sustainability in the context of digitalization and internet use, particularly in AI, which requires large servers for data collection and storage.

Participants stated that AI systems should have an intuitive interface, which is understandable to and usable for people without technical or medical expertise. Adaptable explanations, different languages, and support staff were suggested to ensure comprehensibility for all ages and levels of education. Participants also emphasized that AI systems should be accessible to all population groups.

Participants emphasized the importance of user training and education for the successful application of AI, recommending short instructions and learning platforms for patients who use AI systems independently. They argued that physicians using AI would need training or qualifications, which they should arrange themselves, while others suggested that instructions for new AI systems should be available to physicians, considering integration into medical studies.

According to participants, the development and use of AI could pose a noteworthy financial challenge. Participants questioned whether AI could be made available to everyone for free, and whether health insurance companies would subsidize AI applications.

Participants identified the provision of sufficiently large amounts of recent and representative AI learning and working data as important. They expressed that a large number of training datasets would be necessary for adequate development and would therefore increase acceptance. In some cases, however, data provision was seen as difficult, for example, due to legal or time constraints from authorities or physicians.

Another challenge identified by participants was the need to build trust or acceptance of AI among patients and physicians, through time or positive experiences (see factors promoting acceptance). Participants stated that, as patients, they always had to trust the people treating them first, which would be no different with AI.

A minor aspect mentioned was the challenges faced by health care professionals in integrating AI into their daily work, including the need for additional skills and time to explain and operate the technology. They suggested developing AI with a practical focus, involving users in the process, and gathering feedback to ensure easy integration into daily workflows.

Participants preferred treating AI as a medical device, requiring approval before implementation. It was believed that current evaluations lacked oversight and independence, and suggested alternatives to assess AI systems based on medical benefits, applicability, and ethical considerations, rather than just economic factors. They advocated for transparent quality standards with defined goals and guidelines, and discussed the need for regulatory agencies or legal systems to set these guidelines. They also emphasized the need for oversight institutions monitoring AI operations to ensure compliance and consequences for noncompliance. Minor aspects mentioned were a legislation requiring humans rather than AI to make treatment decisions, as well as the role of government regulation and legal framework in promoting the fair use of AI without stifling development.

The use of AI in medical care was a topic of debate for participants, with opinions ranging from its potential use everywhere, to being relevant only in clinical settings and not for individuals, to being entirely inconceivable. The dominant opinions were imagining AI as a support tool and information source, predominantly for physicians, and skepticism about AI making independent decisions, particularly in medicine, as this could involve health-related decisions and ethical considerations.

Participants mentioned potential future areas of application throughout the treatment process (see the subthemes in Table 2). They considered AI to be beneficial in communicating with individuals who are unable (eg, by suggesting appropriate sentences) or afraid to speak with physicians (eg, due to anxiety or shame-inducing issues) and for adapting language levels or using different languages. Furthermore, AI could be used as a documentation support tool, for instance, during anamnesis or in the creation of drafts for physicians' letters.

Participants suggested that AI could be used in data collection and networking, building a database for physicians, and enabling early intervention through preventive monitoring of patient data. According to the participants, consolidating patient data and facilitating the exchange of expert knowledge could speed up treatment processes, identify correlations, and gain new insights. Furthermore, AI could support patient-physician coordination. For example, in research, AI could overcome language barriers to compile results for global studies.

Diagnostics was identified as a major field, as AI could accelerate the process through objective data analysis and collection, providing a basis for decision-making for physicians. Participants suggested including patient values in symptom checkers to improve diagnostic suggestions’ trustworthiness and reduce uncertainty. AI’s potential in image recognition, processing, and editing was expected to be leading in the near future (eg, in radiology, in neurology—evaluation of EEG data, in dermatology—screening of naevi), due to its already powerful capabilities in this area.

According to participants, AI could be used in therapy to help ensure that side effects, drug interactions, and current knowledge are taken into account. AI could be useful in providing rehabilitation and physiotherapy support by creating appropriate exercise plans with motivational elements, individual adaptation, and monitoring of exercises. Another possible area was operations, where a personal or verbal component seemed less important.

Participants discussed the use of AI in care and entertainment to address the shortage of skilled workers. Robots (featuring AI) could simplify work for carers and provide companionship for those in need of care. Beneficially, robotic animals would not cause allergies and would not have any physical needs. However, it was argued that AI alone should not be the answer; rather, it should be a change in approach to care to ensure enough carers.

Another area of application identified by participants was process management. They found AI useful in medication provision, operating theatre, and bed management, and patient triage. They identified great potential in administrative tasks, like optimizing ordering systems with faster appointment allocation, assisting patients with follow-up by providing necessary information and reminders, or physical assistance with luggage robots.

Participants’ opinions differed regarding the potential use of AI in care and invasive interventions, ranging from possible (see above) to unthinkable. They expressed a lack of confidence in AI’s ability to perform operations reliably, as there would be no error tolerance. They could not imagine flexibility and short-term adaptation times (which would be required, for example, due to individual anatomy or the occurrence of errors) in the AI. Participants opposed AI in care, arguing that human interaction in care is crucial and should not be replaced by AI. They feared that vulnerable people in need of care (such as people with disabilities, children, and older adults) could be further excluded from society through the lack of human contact. There were comments about finding the idea of using care robots or animal robots sad, questioning how society will deal with people who need support in the future. In addition, operating these devices could also cost carers more time, potentially further reducing human contact.

Other areas that respondents felt were unsuitable for AI were tasks requiring empathic conversations, certain specialties, and sole AI use in therapy, regardless of specialty. Participants expressed distrust in AI’s ability to possess empathy and understanding of the human psyche (and its illnesses), which was mentioned as especially important in conversations. They also questioned whether AI could make appropriate therapy recommendations and provide support during difficult times. Participants highlighted that they would not want to be informed by AI about serious illnesses, or would be unsure how to deal with such a situation, because of the need for human contact in these settings. Participants opposed AI in gynecology or urology, either due to the sensitivity of health matters shared or for other unspecified reasons. AI in general medicine was also perceived as inappropriate, as patients often seek personal contact. A minor aspect was that research could not be imagined as a potential area for AI, as it requires human foresight. Subthemes with descriptions of thinkable and unthinkable tasks provided by AI are presented in Table 2, while participants’ quotes are presented in the Multimedia Appendix 1.

The UTAUT can help identify factors influencing user acceptance, especially among those hesitant to adopt new technologies, and can be applied in the development of technical innovations [42]. The results of our study regarding patients’ acceptance partly align with the main determinants of UTAUT (performance and effort expectancy, social influence, and facilitating conditions).

Participants expected AI to improve care or living conditions (performance expectancy). Important acceptance criteria included simple functionality and handling, and easy-to-understand explanations (effort expectancy), which were in line with previous studies [32,53]. Knowledge about AI can increase its acceptance [32,54], and both our participants and other stakeholders, including the European Commission, consider the transfer of knowledge and the training of medical staff and users to be challenging but indispensable prerequisites [55-59].

The participants valued recommendations from peers and physicians (social influences), stating negative attitudes from physicians would reduce their acceptance [60]. In line with the literature [61], participants’ acceptance would increase if AI was tested in studies and in practice, but no one wanted to be the first to test the system.

Factors promoting participants’ acceptance (facilitating conditions) were in line with previous studies and included data protection, patient- and nonprofit-oriented development and implementation [62,63] and a large and representative database for AI [63-65]. Previous studies have demonstrated that data protection and transparency in data use are essential for development, trust, and acceptance of AI [62,63], which is consistent with our findings. To prevent data leaks, robust security measures must be implemented, but this can hinder the acquisition of a sufficiently expansive and representative database [66]. In accordance with the literature, our participants viewed the attainment as a challenge, but an imperative requirement for AI to function adequately [63-65]. Furthermore, participants and current literature also discuss that AI systems should be trained on diverse data [67] to avoid inheriting existing inequalities from models or the training dataset [65]. Due to the sensitivity of health care data, participants emphasized its protection [68] and trustworthiness of entities receiving their data, which would be essential for data sharing [69]. Trustworthy AI should be in compliance with ethical and legal regulations as well as technical and social functioning throughout its lifecycle [70]. The US FDA, which approves medical AI devices, recommends cybersecurity, risk management, as well as postimplementation monitoring and evaluation, and calls for adaptive, science-based regulations that protect against risks without limiting benefits [71]. The European AI Act identifies AI in health care as a high-risk application, as it deals with personal data, and also emphasizes transparency, cybersecurity, and risk management throughout the AI lifecycle, and data governance to ensure representative and error-free training data [72]. Patients strongly opposed selling health data to private companies for AI research, despite some arguing it could be justified if the product benefits patients [73]. This distrust toward private health care companies was also evident in previous studies [74-76]. One of the most important findings was that participants desire independent AI development and funding aimed at patient benefit, with concerns being justified as AI may be used to increase profits [77]. The FDA also sees this risk, although it acknowledges that the relationship between financial optimization and improved health outcomes can be complex and result in financial disadvantages for provider organizations, insurance companies, or health systems. However, sponsors should be transparent and focus on health outcomes, and a comprehensive and regular approach across the health system is needed to counter the negative risks of financing and keep pace with the development of AI [71].

The UTAUT model provides a good orientation for measuring acceptance and intended use of AI. Due to the models’ criticized lack of complexity [43,78], UTAUT (and TAM) have been adapted and extended for use in the health care sector [79,80]. Nevertheless, it is only partially applicable to all health care issues and their stakeholders, including patients, physicians, and carers. Thus, as our results also show, sociocultural aspects or factors such as training or integration into everyday working life are important in health care applications [43,81] and need to be integrated into these models to reflect the complexity of digital applications in health care [45].

Participants emphasized that AI should be accessible and usable by all people, not exacerbating inequalities, as could be the case through biased training data, as mentioned above, or a lack of technical or financial possibilities [82].

In the context of resource use, the issue was mentioned that environmental sustainability should be considered when developing new AI systems. In a Swedish survey of health care managers, the climate aspect was mentioned in connection with the successful implementation of AI [58], and the European Commission identifies environmental sustainability throughout the lifecycle of AI as a requirement for trustworthy AI [56]. There are also efforts to assess the environmental compatibility of AI [83] and to identify the environmental impact of medical digitalization [84]. Participants considered environmental sustainability necessary, as a global digital infrastructure already consumes many resources and has a high carbon footprint. The participants' assumption is not unfounded: in 2019, 3.8 % of greenhouse gas emissions were attributed to the digital sector [85,86], and the trend is rising. AI has great potential to promote sustainable development and reduce the environmental footprint [4,87]. Yet it also has negative environmental impacts that require careful use and a balanced approach involving regulation and all stakeholders [88-90].

Comparison of our findings with existing literature confirms that patients prefer AI applications to be certified by external, independent institutions [62]. Participants and other health care stakeholders agreed that AI should meet quality standards, similar to medical devices [8,82]. As noted above, a regulatory framework was seen as crucial to the development and implementation of AI [58,65,91], guiding the development process without hindering progress and requiring oversight during application [65]. Assuming that AI could be fed with a global dataset in the future, then globally applicable regulations would be a logical consequence, although their implementation would undoubtedly prove challenging. Although there are recent developments, such as the European AI Act [72] or American regulatory approaches to AI in medical applications [71,91,92], trying to introduce compatible standards, there seems to be a lack of global guidelines.

Despite participants expressing a general openness toward AI, there is considerable skepticism, even among tech-savvy participants, about AI-based decision-making and conversational guidance [73]. Particularly in the case of patients with chronic or terminal illnesses, the human factor appears to be of paramount importance [32,93]. Thus, some participants, who could imagine using AI for surgery, were against its independent use, but in favor of AI support [31]. In accordance with literature [94], opinions ranged from the independent use of AI for minor interventions to no use at all in surgical procedures.

As a key finding of our study and in line with previous studies, participants preferred AI as support systems with human supervision rather than autonomous systems [59,62,95]. As the lack of human involvement is a relevant barrier to acceptance of AI systems [61,95], potential applications should focus on support rather than independent functioning.

Our study indicates that most participants can imagine AI being used in diagnostics or data processing [96]. In line with literature [30,95], participants attributed a more objective diagnostic capability to AI and felt that AI could quicken the diagnostic process [30,60] and assist physicians, including in the identification of rare diseases [95].

AI-based systems already exist for outpatient and inpatient care [97]. Strikingly, there were substantial differences in participants' opinions regarding the use of AI in care. People in need of care may no longer be able to advocate for themselves, which increases the need for protection and raises ethical questions about the use of AI in care [97]. Furthermore, care is an intimate setting that involves both physical and emotional aspects, where maintaining communication and fostering a trusting relationship are of great importance [96]. Many participants felt that these aspects could not be provided by AI and therefore considered its use in care undesirable for activities beyond the manual relief of staff. In line with Deckert et al [96], they concluded that AI could only be integrated into care to a limited extent and could not replace nursing staff. Maintaining the interpersonal dimension would be a challenge [97] that skeptical participants felt AI could not meet. Other participants expressed the opposite view, arguing for the integration of AI in care, often stating that AI would be better than no contact at all. The diversity of opinions and concerns underscores the importance of a balanced approach to implementation in care that combines AI and humanity [97]. Furthermore, participants could not envisage AI in psychology, where conversational interactions are central. Nevertheless, some stated that certain patients might find it easier to confide in an AI than in medical staff concerning shameful issues or fears. Szalai [98] describes this aspect in the context of borderline therapy. Furthermore, conversational AI is advancing, as it is capable of engaging in a moderate conversation through language processing, using psychotherapeutic techniques [99]. Despite the existence of numerous potential applications for AI in mental health treatment, its actual use in clinical practice remains limited [97,100]. Most machine learning solutions for mental health are developed without involving end users and their individual needs, which can create barriers to using existing options [100]. Furthermore, there are concerns that human characteristics such as imperfection [101] or empathy [59] are necessary for psychological treatment, a perception shared by our participants. Exclusively technically generated treatment plans do not include the complete assessment by and emotional awareness of physicians [102]. This can critically reduce treatment success and discourage patients from continuing treatment [103]. AI has the potential to support mental health care, which seems particularly beneficial in light of the growing need for it [104]. However, in this specialist area, it seems particularly important to maintain basic ethical principles and physician involvement [104].

As we examined patients’ perceptions in two regions of Germany (south-west and central-east), and existing German studies in other health care settings have reported similar results [32-34], our findings are applicable to Germany as a whole. Patients’ perceptions of AI are also similar in comparison with other European countries [31,61] and industrialized nations [29,60].