Advances in machine learning and image recognition have driven the implementation of systems based on artificial intelligence (AI) in clinical practice. These systems are able to perform tasks commonly associated with intelligent beings, such as reasoning or learning from experience [1]. AI systems based on machine learning gain insight directly from large sets of data [2]. Such AI systems can, therefore, make more complex decisions and complete more complex tasks than can rule-based systems. Moreover, machine learning has the potential to improve AI system capabilities with growing amounts of training data. In medical disciplines such as radiology, AI systems already diagnose specific diseases in a manner comparable with that of experienced physicians [3]. Current AI systems provide diagnosis and treatment recommendations, facilitate access to information, and perform time-consuming routine tasks of physicians, such as the segmentation of radiological images [4]. In the light of an accelerated technological development, AI systems in health care are expected to pave the way for truly personalized medicine by augmenting medical decisions and helping physicians cope with increasing amounts of relevant data and growing numbers of medical guidelines [5]. In fact, future AI systems are expected to autonomously execute medical actions (eg, communicating results and making diagnosis decisions) as efficiently and accurately as expert physicians [3]. Thus, medical professionals will be able to delegate more complex medical tasks to those systems than to any traditional rule-based clinical decision support system [6].

Despite their benefits and the great expectations toward their future use, AI systems do not cause only positive reactions in medical professionals. Although nearly all current applications of AI are still limited to narrow use cases, the future potential of this technology has resulted in a discourse driven by a duality of hyped expectations and great fears [7]. Many physicians seem to be concerned about the changes that AI systems impose on their profession [8]. Such negative attitudes toward new technology can manifest as resistance attitudes, resulting in hesitation toward adopting a technology [9] and even in active resentment against using a technology in clinical practice [10]. Although multiple studies have investigated attitudes toward AI [11-13], they did not consider how resistance attitudes toward AI are formed. Specifically, they did not examine whether negative attitudes toward AI stem from the perception that AI is threatening the medical professional identity. Therefore, we aim to address the following research questions:

Professional identity refers to how professionals define themselves in terms of their work roles and provides an answer to the question “Who am I as a professional?” [14]. In general, professionals strive to maintain a coherent and positive picture of themselves [15], resulting in a tendency to interpret experiences in identity-coherent ways. This helps individuals to adapt to social changes, such as technological innovations, and to create benefits from the positive identification with a social group. Professional identity can be considered as a combination of social identity [16] (ie, membership in a group of professionals) and personal identity (ie, the individual identification and enactment of professional roles) [15]. Medical professionals are known for their strong commitment to their professional identity, which is already developed in the early phases of their socialization and later refined through practical experience [17]. The medical profession’s group boundaries are very rigid and shaped by strong core values and ideals [18]. Medical professionals can, therefore, be seen as a prototypical profession [17], and their professional identity is particularly resilient to change [19].

Identity threats are “experiences appraised as indicating potential harm to the value, meanings, or enactment of an identity” [20]. Such experiences are potentially decreasing an identity’s value, eliminating a part of the identity, or disturbing the connection between an identity and the meaning the individual associates with the threatened identity [20]. In the following, we distinguish between two parts of individuals’ identity that can be threatened: personal identity and professional identity.

First, self-threat describes a context-independent threat to personal identity by challenging fundamental self-motives of distinctiveness, continuity, generalized self-efficacy, and self-esteem [21]. Self-threat can bias information processing, result in avoidance of threatening information [22-24], and adverse emotional reactions [25]. Recently, self-threat has been identified as an antecedent of resistance to technology [26].

Second, medical professionals can be threatened along different dimensions of their professional identity. Drawing on a synthesis of prior work (Multimedia Appendix 1 [9,10,14,27-34]), we differentiate between 2 dimensions of medical professional identity threats. First, threats to professional recognition refer to challenges to the expertise and status position of medical professionals [10,27]. Second, threats to professional capabilities refer to the enactment of roles related to the medical work itself. The latter include threats to the care provider role [28], autonomy [14,29-31], and professional control [9,32]. Multiple studies show that professional identity threats can stipulate resistance to new technologies and organizational change [10,28,35]; however, it is unclear how these threats manifest in the context of AI systems.

Multiple conditions do influence how medical professionals experience identity threat from AI systems. In this paper, we focus on the following two conditions: professional experience and the perceived temporal distance of AI. First, perceived experience influences how likely medical professionals perceive that AI systems can replace parts of their work. In particular, more experienced physicians believe that they have unique skills that AI systems cannot substitute, whereas novices have yet to develop those skills. Second, medical professionals might have different perceptions of how fast AI systems are implemented in clinical practice. Perceiving AI systems as temporally close and relevant in the immediate future suggests that AI systems will be seen as more influential on concrete medical work practices, thus threatening one’s professional capabilities. Conversely, if AI systems are perceived to be relevant only in the distant future, the perceived threat might be less specific to medical work practices but more relevant for the long-term reputation of medical professionals, thus threatening their professional recognition. Hence, the perceived temporal distance of AI systems could affect how relevant each dimension of the professional identity becomes [36].

We collected data in 2 waves of a web-based survey. The first wave survey mainly addressed medical students, whereas the second survey focused on experienced physicians from different specialties. All participants were provided with a vignette of an AI system named Sherlock (Textbox 1) that was based on the description of IBM Watson and was pretested with researchers and medical students. We selected this vignette because it depicts a general AI system in which the participants were familiar with because of the marketing efforts of the vendor. It does not limit the abilities of the AI to a specific medical specialty. The vignette was purposefully focused on the benefits of the system and evoked expectations of high accuracy to establish the picture of a powerful AI system that goes beyond extant narrow use cases and, thus, has the potential to be threatening. Then, control questions were included to ensure that participants associated AI with the vignette and that it was perceived as realistic. The vignette was pretested with medical students and professionals. We then asked an open question about the participants’ perceptions of the changes to their professional role caused by AI systems to gain qualitative insights into the perceived upcoming change of their identity. Afterward, we asked participants to complete the provided survey of experience, identity threat, resistance attitudes, and perceived temporal distance of AI systems.

We used a 4-item measure of self-threat [21] on a 6-point Likert scale and a 3-item measure of resistance [9]. Furthermore, we asked participants whether they perceived the change from AI systems as temporally close or distant by assessing the agreement to the following statements: “Such systems will only become relevant in the distant future,” “Such systems are unlikely to be implemented technically,” and “Such systems are too abstract and intangible for me.” We extended existing measures for threats toward professional recognition and professional capabilities following the procedure of MacKenzie et al [37], as outlined in Multimedia Appendix 2 [9,28,29,37-46]. For medical students, we also assessed positive expectations toward AI, which mirrored the negatively framed items that we used for identity threat (identity enhancement). We also included an open question about their expectations of how the medical role would change with the introduction of AI systems. As a control variable, we asked for participants’ familiarity with clinical decision support systems. Except items related to self-threat, all items were measured on a 5-point Likert scale from totally disagree to totally agree. Table 1 lists the survey items. The items for identity enhancement can be found in Multimedia Appendix 3 [47-50].

We validated the scales of professional identity threats in the sample of novices and experienced physicians and the corresponding identity enhancement values, as outlined in Multimedia Appendix 2. A confirmatory factor analysis with all measurement scales resulted in a good model fit. All scales displayed good psychometric properties, including reliability, convergent validity, and discriminant validity (Multimedia Appendix 2). The correlation between threats to professional recognition and professional capabilities was 0.64, which was smaller than the lowest square root of the average variance extracted of 0.77, indicating acceptable multicollinearity. Similarly, multicollinearity between self-threat and threats to professional recognition was acceptable with a correlation of 0.66 being lower than the lowest square root of the average variance extracted of self-threat. We accounted for potential common method bias in the survey design and through testing for a common method factor. The results indicated that common method bias is unlikely to have a strong impact on our results (see Multimedia Appendix 2 for details). The items of identity enhancement were combined into 1 factor because of the result of the exploratory factor analysis. All analyses were performed using SPSS (version 26; IBM Corporation) and Stata (version 16; StataCorp).

A total sample of 227 novice and experienced physicians participated between fall 2017 and spring 2019. Participants were recruited from medical social media groups and by personal reference. After excluding participants because of failed comprehension checks or poor data quality (ie, very fast completion time or answers to the open question that were completely unrelated to the question), a total data sample of 206 participants was used for data analysis. Of these 206 participants, 164 (79.6%) were medical students and 42 (20.4%) participants were medical professionals across different specialties (see Table 2 for details on sample). We included both medical students and trained physicians in our sample for two reasons: First, especially inexperienced physicians are susceptible to the influence of technology [51]. They may thus provide valuable insight into the effects of AI systems. Second, particularly medical students face strong, long-term career consequences if AI systems alter the meaning of specific medical disciplines such as radiology. They are thus likely to cognitively engage with potential identity threats to make reasonable career decisions, for example, with regard to the specialty they pursue. Conversely, experienced medical professionals may have a more pronounced professional identity and may experience threats from AI systems differently.

Our main outcome variable measures participants’ resistance toward using the Sherlock application. Because of differences in sample size between novice physicians (n=164) and experienced physicians (n=42), we conducted a nonparametric Mann–Whitney U test to test the differences between the 2 samples regarding resistance attitude and self-threat. We conducted seemingly unrelated regression analyses with self-threat and resistance as the dependent variables because we expected their error terms to be correlated. The predictors were included in a stepwise approach to assess how much additional variance they explain. To test the effect of perceived temporal distance of AI, we incorporated 2 interaction terms between temporal distance and the 2 dimensions of professional identity threats into the regression analysis. We conducted a mediation analysis following Hayes [52] with 10,000 bootstrap samples to test the mediating role of self-threat on resistance attitudes.

Ethics approval was not sought or obtained from the Ethics Committee at the University of Mannheim, as an approval was not required under the institution’s current ethics statute.

We find that both experienced and novice physicians perceived identity threats from the upcoming change from AI systems (Table 3). Novice physicians showed relatively high resistance to AI and self-threat, whereas experienced physicians showed slightly lower resistance and self-threat from AI. The group differences were significant for resistance (P=.005) and self-threat (P<.001). Novices perceived equally strong threats to their professional recognition (mean 3.05, SD 1.23) and professional capabilities (mean 3.25, SD 0.97), whereas experienced physicians perceived a stronger threat to their professional capabilities (mean 2.72, SD 1.18) than to their professional recognition (mean 2.38, SD 1.03). Group differences were statistically significant; novices experienced more threats to professional recognition (P<.001) and professional capabilities (P=.80) than experienced physicians. Similarly, novices reported AI systems as slightly more temporally distant than did experienced physicians (P=.08). Moreover, in the sample of experienced physicians, the experienced threats and resistance attitudes differed based on the medical specialty (Table 4). The descriptive statistics show, for example, that physicians in the psychiatry specialty reported stronger threats to professional recognition than to professional capabilities, whereas surgeons reported stronger threats to professional capabilities than to professional recognition. However, as the sample size was relatively small, more research is needed to fully understand the effects of different specialties on experienced identity threats.

Testing the relationships between different types of identity threat and resistance attitudes in the total sample (Tables 5 and 6; Multimedia Appendix 3), we found that perceived professional identity threats directly affected resistance attitudes and personal identity threat (self-threat). Both threats to professional recognition (P<.001) and threats to professional capabilities (P<.001) were significant predictors of self-threat (model 4a, Table 6). Moreover, we found that both professional identity threats contributed independently to resistance to change. However, threats to professional recognition predicted resistance only in isolation (Multimedia Appendix 2; P<.001) but not in combination with threats to professional capabilities (Multimedia Appendix 3, model 3b; P=.50). Hence, threats to professional capabilities overruled the impact of threats to professional recognition on resistance attitudes and significantly increased resistance (Multimedia Appendix 3, model 3b; P<.001). The findings suggest that threats to professional recognition are more strongly related to personal identity, whereas threats to professional capabilities are more strongly and directly related to resistance to change.

We also analyzed how perceiving AI systems as temporally close or distant interacted with perceived professional identity threat. The perception of AI systems as temporally distant interacted positively with threats to professional recognition (P<.001) and interacted negatively with threats to professional capabilities (P=.02) in predicting self-threat (Table 6, model 4a). In predicting resistance, the perception of AI systems as temporally distant interacted negatively with threats to professional capabilities (Table 6, model 4b; P=.04), whereas the interaction with threats to professional recognition was not significant (P=.24). Figures 1-4 show the moderating effects of temporal distance on both dimensions of identity threat. The findings suggest that experienced identity threats are closely related to how temporally distant or close the technological change from AI systems is perceived. Threats to professional capabilities refer to more concrete and context-specific elements of professional identity. Thus, these threats are more salient if physicians believe that AI systems are temporally close and relevant to clinical practice in the near future. Conversely, threats to professional recognition require physicians to consider their profession in a holistic way. Thus, these threats are more salient if physicians perceive AI systems to be relevant only in the distant future.

For the sample of novice physicians, we also collected data about perceived identity enhancements through AI. We used these data for robustness analysis to exclude the possibility that our results regarding the effects of identity threat are biased by hidden positive attitudes. The identity enhancement items mirrored the wording of the identity threat items to capture any positive expectations toward the change induced by AI systems (Multimedia Appendix 3). The analysis shows that identity enhancement reduced resistance to AI (P=.02) and was not related to self-threat (P=.52). After including identity enhancement as a control variable, the above-described effects of perceived identity threat remained qualitatively unchanged. This indicates that perceived professional identity threats have a significant effect on resistance to AI over and beyond any effects of perceived identity enhancement through AI.

Finally, we conducted a content analysis of the qualitative statements from novice physicians to validate that our measured dimensions capture the experienced changes through AI systems. The data set consisted of a total of 414 distinct statements by 176 participants. The content of all statements was classified as positive or negative statements about AI systems by 2 independent coders (EJ and one student assistant; Table 7).

One-third of the negative statements (34/105; 32.4%) described threats to professional recognition. The implementation of AI systems was perceived as leading to a loss of status and prestige for the occupational group of physicians and made participants fear that physicians might become redundant and reduced to a mere voice of the AI system. Moreover, participants feared that expert knowledge would become less important as AI systems incorporated more up-to-date knowledge than ever possible for a human being. The statements also contained multiple threats to professional capabilities. As such, participants feared that physicians might lose their autonomy in decision-making as they might trust the AI system more than appropriate, whereas the system would perform tasks autonomously. In addition, participants perceived that it would become more difficult to be a care provider with AI systems in place, as these systems would increase the distance between physicians and patients. The participants also feared that liability issues would arise if they disagreed with AI decisions. Conversely, 3 categories of positive statements emerged from the content analysis. AI systems were perceived as supporting decision-making through reduced uncertainty and complexity in diagnostics. Moreover, they were seen as facilitators of access to knowledge, supporting especially novice physicians, by providing access to the newest guidelines and empirical findings.

Our work contributes to the knowledge on the impact of AI and the future of work in health care [53-55]. It shows that professional identity threats from AI systems are indeed a serious concern for novice and experienced physicians and contribute to resistance to AI. AI systems threaten both professional recognition and professional capabilities of medical professionals. Threats to professional capabilities directly contribute to resistance to AI, whereas the effect of threats to professional recognition is mediated through self-threat. Professional experience and perceived temporal distance of AI systems influence the relationship between perceived identity threats and resistance attitudes. Medical novices experience stronger identity threats than medical professionals. In addition, if AI systems are perceived as more relevant in the near future, threats to professional capabilities are more profound. If, however, AI systems are perceived as relevant in the distant future, threats to personal recognition gain in importance.

Our findings have implications for the understanding of how the medical professional identity changes with increasingly powerful AI systems and how AI systems are integrated into medical decisions. First, experienced identity threats influence how physicians adapt their professional identity to the upcoming change. For instance, study participants who indicated that “the role of the physician will be more passive, since decisions will be automated” might be less likely to choose specialties such as radiology. This can lead to fewer physicians who actively work with AI systems and develop the technological capabilities to evaluate those systems. Furthermore, several participants declared that they planned to focus on soft skills instead of analytical decision-making skills, which would rather be performed by an AI. Thus, instead of using AI systems as a second opinion and engaging in elaborate decision-making, physicians might end up delegating important tasks to AI systems without considering them in detail.

Second, threats to the professional identity cause identity protection responses [20] that directly impact technology use. In health care, physicians are pivotal for developing the ground truth for learning algorithms and for identifying relevant explanations and predictive values. Furthermore, physicians can make better diagnosis decisions with the support of trained algorithms and use them as a second opinion [56]. However, if they feel threatened in their identity, physicians are less likely to engage in the active development and adaptation of AI systems and resist their implementation. Moreover, identity protection responses can lead to incorrect medical decisions with AI systems if physicians reject AI advice as soon as it contradicts their opinion and is perceived as threatening [51]. In particular, threats to professional capabilities play a focal role in developing negative attitudes toward AI systems and should, thus, be addressed through specific medical training in interacting with AI systems.

This study has several limitations that can serve as a springboard for future research. First, by using the survey method, we were not able to capture how the identity develops through a longer period of time and whether medical students who perceived stronger threats to their future from AI would switch to a nonthreatened profession that requires more subjective interpersonal skills. In addition, it would be interesting to see how the professional identity is affected in clinical practice through a more intensive interaction with AI systems. Furthermore, our study provides first insights into potential differences in experienced identity threats across medical specialties. Specialties such as radiology or pathology were scarce in our sample, although those specialties often use AI in medical practice. Consequently, a follow-up study that looks at differences across specialties in more detail might provide interesting insights. In addition, our sample consisted of respondents who reported a relatively low degree of familiarity with AI systems. This reflects the current situation in medical education, in which medical novices are not trained in the use of AI systems. However, whether a sample with more familiarity would experience lower degrees of threat from AI systems needs to be further researched. Second, as noted in the literature [28,33,57] and underlined by the qualitative survey responses, there are also positive appraisals of AI systems that can enhance, rather than threaten, individuals’ identity. Given that there are both strong positive and negative perceptions of the impact of AI systems on the professional identity, future research should consider the impact of ambivalence [58] on professional identity formation and restructuring. Third, we presented an AI system with a 90% accuracy rate. However, in clinical practice, the accuracy rate is highly dependent on the context, that is, the complexity of patient cases, and can be heavily disputed by medical professionals. Furthermore, with lower perceived or actual accuracy, physicians might develop more negative attitudes toward the AI system. Finally, as our study examined only 2 dependent variables, it is important to investigate how professional identity threat from AI systems impacts other variables, such as anxiety and long-term behaviors.