Perceived credibility in human-AI communication for medical information: mapping a choice mindset surrounding algorithm authorship and recommendation

Introduction

The term one-to-one human-machine communication (HMC) concisely depicts at least part of the quintessence of smart healthcare, or personalized digital healthcare, characterizing the age of “Healthcare 4.0”—for an industrial review, refer to Jayaraman et al. (1). In previous one-to-one HMC literature, while a body of research centers on developing accurate, fair, robust, explainable, transparent, inclusive, empowering, and beneficial smart systems powered by artificial intelligence (AI) (2-4), the human-centered tradition has pointed to the critical role of the user’s choice determining to what extent such smart systems can be accepted and utilized. As a research topic under smart system, automated authorship, characterized by algorithms that require no human intervention after initial programming, has given rise to a scholarly focus on how users evaluate information not generated by human experts. In this context, the psychological constructs for decision-making based on perception and evaluation, like choice mindset (5), could provide a user-centered approach for AI user research. Meanwhile, although many prior studies have touched upon credibility perceptions in human-AI communication, broaching topics like user trust in AI recommendation systems (6) or acceptance of healthcare chatbots (7), virtually no existing conceptual work has mapped the interplay of human-AI communication dynamics in medical information seeking, particularly choosing between algorithm authorship and recommendation based on perceived credibility of these two information sources, across health-related contexts. Specifically, the notion of algorithm authorship here refers to content creation by an AI system, as mainly addressed in AI-generated content (AIGC) discourse, whereas algorithm recommendation refers to content selection or endorsement by an AI system from a set of existing options, as discussed in recommender system literature (8).

As the instrumental rationale for the present conceptual paper, the choice mindset refers to the state of mind in which people perceive the act of selecting one’s behavioral option among several through a lens of choice (5), which influences individual decision-making and interactions with others, like negotiations (9). Notably, while the choice mindset helps increase self-control in personal decision-making, it can also increase self-centeredness in relationships with others, reducing trust and cooperation (10). The notion of “others” here—if interpreted broadly as socialization agents or social actors—may extend to AI in the HMC context, which has been indefatigably addressed by the advocates of the Computers-Are-Social-Actors (CASA) paradigm—for a discussion on this paradigm, see Lee et al.’s study (11). From a user-centered perspective, when a personal task requires assistance, how would a user make a choice between a human helper and a machine agent, based on trust, justice, and/or any other concerns? To address this, mapping and advancing the key concepts for navigating empirical investigations could be a starting point. Advancing a theoretical synthesis and concept map regarding credibility issues in smart healthcare user research, the current research aims for unpacking the source effects of algorithm authorship and recommendation on perceived credibility in one-to-one human-AI communication for medical information, while foregrounding these two types of AI expertise that could be crucial for developing transparent and explainable AI.

Theoretical synthesis

Today, it is not uncommon for people to turn to a chatbot for medical information addressing personal health concerns. Conceptually, guided by the choice mindset, the present article attempts a theoretical synthesis for one-to-one human-AI interaction in such medical information seeking, covering not only the upstream evaluation as an extensively studied field with renowned constructs like Elaboration Likelihood Model (ELM)—for a review, see Petty et al.’s study (12)—but also the downstream decision-making. Extending this concern, an emphasis is placed on the perceived credibility of AI agents and the medical information generated by them in human-AI communication, as a central issue of smart healthcare. In recent clinical practices, the issues surrounding perceived credibility in concrete, specialty-specific applications of AI have also earned increasing attention. For example, large language models (LLMs) like GPT-4 and Claude-3-Opus have been employed in simplifying interventional radiology reports to enhance patient comprehension, showing high performance in clarity, completeness, and actionability; yet these models also exhibit trust-breaking errors, which could diminish patients’ confidence (13). Looking beyond clinical contexts, our focus is placed on credibility dilemma in AI-powered patient-facing medical communication, which requires our framework of choice mindset to address algorithmic trust in daily life settings.

As a research topic under HMC, automated content generation, characterized by algorithms that require no human intervention after initial programming, has given rise to a scholarly focus on how users evaluate information not generated by human experts. Around the relationship between automated content generations and user perceptions, mixed results have been presented (14,15). A line of research reveals that algorithm-generated contents are overwhelmingly preferred by readers over manmade ones, behind which credibility judgments play an important part and can be explained by different machine heuristics. For example, You and colleagues (14) confirmed via a comparison of algorithmic vs. human suggestions that individuals showed appreciation for algorithms to a large extent, and that users were more likely to accept algorithmic suggestions than human versions. Sundar (16) posits that modality, agency, interactivity, and navigability constitute the cues for cognitive heuristics and thereby affect credibility evaluation; these constructs, mainly based on the features of digital media technology, can be applied to the investigation of algorithm-generated advice, shedding light on whether machine heuristics affect users’ perceptions of contents attributed to either an algorithm or a combined authorship—yet, evidence regarding the latter remains scarce (17). Grouping relevant topics and concepts, what follows is a selective review on the underpins of credibility in AI-based persuasion, and specifically in AI-based health advising. Methodologically, this approach allows for an integrative synthesis across disciplines, yet without offering empirical validation. It is thus necessary to also note that this conceptual paper is inherently shaped by our selection of existing literature, which can use separate discussion on the ethical considerations of conducting reviews involving sensitive medical technologies, particularly in relation to health equity, risk of amplification of existing bias, and responsibility in framing AI credibility for vulnerable populations. In this regard, while moving forward to clarify the crucial concepts for our mapping work, we encourage future reviews to systematically address the potential limitations left behind here, such as selection bias and the interpretive nature of our theoretical synthesis.

Perceived credibility in AI-based persuasion: expertise and trustworthiness

Usually regarded as a communication variable in persuasion research, the credibility of a source influences the interaction between the communicator and the message recipient (18). Not as an intrinsic property of the source, notably, credibility refers to the subjective judgment regarding the source’s believability, therefore varying among individuals. In this view, perceived credibility, as addressed in this paper, could be a more precise description than the notion of credibility alone; only for simplicity, credibility remains in use hereafter. In existing studies on credibility’s effects, two primary dimensions, expertise and trustworthiness, are usually not separately manipulated, while they are conceptually distinct, as supported by factor analytic findings (19). Therefore, separate attention will be given to the two aspects of credibility in the following discussion, especially taking the context of AI-based persuasion into account.

Expertise is considered a reflection of knowledge or ability ascribed to the communicator in a particular field (18). Recent advances in AI technology have convinced more people about the seemingly limitless capabilities of AI than ever, regardless of the specialty of interest. In this sense, it seems natural for a user to presume that an AI-powered expert system should be convincing. Typical examples can be easily found: many students today turn to AI bots for scoring and/or rephrasing their essays, with a blind belief that AI should be an expert in writing; when searching for references, some students would like to accept AI recommendations as advice from experts—for commentary on chatbot-supported thesis writing (20). It is apparent that heuristics about AI expertise can be easily found in both AI authorship and AI recommendation. However, the two kinds of AI-generated contents, articles and recommendations, do not always achieve their influence through the same psychological processes in every situation, especially for persuasion drawing on expert appeal. To the best of our knowledge, previous literature has yet to systematically differentiate and compare algorithm authorship and recommendation when examining the source effect of AI expertise, whereas separate discussions on each expertise type can be found, mostly spotlighting their differences from human counterparts: for example, see Tandoc et al.’s study (17) for authorship, Wien et al.’s study (21) for recommendation. On the other hand, if given a truth, whether a communicator is likely to tell the truth gives rise to the issues of trustworthiness, which depends on the perceptions of the communicator’s honesty, character, and safety (18). Translating the definition of trust in interpersonal communication into the human-computer communication version, people’s trust in computers has been approached anthropomorphically (11,22). In this view, an AI persuader can be trusted only when some trust-inducing human characteristics are perceived. This also implies the need for considering a wide range of concerns that may determine the level of trust, two of which will be included in the later section further discussing AI trustworthiness.

Aside from the two aspects of credibility above, persuasion researchers have naturally placed considerable attention on the question of how various characteristics of the communicator influence the outcomes of the persuasive efforts. Extending to AI-based persuasion, a gamut of source factors has been examined and reapproached with AI being the communicator (23). The current paper merely focuses on one communicator characteristic, i.e., credibility, while leaving behind other source factors, such as AI’s likability and similarity to the users. As for the functioning of AI credibility in persuasion, one general pattern seems tempting: in line with the ELM, when an issue becomes more personally relevant to a receiver, the source’s expertise and trustworthiness would become less important, as the message contents (e.g., arguments, recommended items) per se earn comparatively greater importance. Similarly, as maintained by the choice mindset, it depends on the salience of choice to activate an analytic mindset (10). Placed together, when receivers find a topic personally relevant and important, they are more likely to engage in systematic processing of the AI-generated contents, or AIGCs, and minimize reliance on peripheral cues including those related to AI credibility. However, this account fails to explain why under some conditions of high personal relevance, source credibility may be considered as the starting point of evaluation. A case in point is: if without any prior knowledge on a certain topic of importance, one must first decide whether to accept a source as credible, simply because, for one reason, this person has no confidence in directly evaluating the contents, which is a classic layman’s dilemma. This limiting condition leads us to an extended discussion delving into the context of smart healthcare.

Other concerns over AI trustworthiness in smart healthcare: privacy and caring

It should be clarified that the specific type of AI falling into the scope of the current paper is symbolic AI, rather than embodied AI. Symbolic AI, as maintained by Franklin (24), refers to AI that functions mainly by managing symbols (e.g., texts, images) within computers; being embodied, on the other hand, means that an AI system could go beyond the computing task and take actions in the real world. While the smart healthcare practices include both types, only the former is addressed in the present research. Though not beyond reproach, symbolic AI is replacing many human experts as alternative service providers. Over the past decades, symbolic AI, represented by chatbot technology, has become a burgeoning field, with practical applications in many sectors such as banking, customer service, medicine, education, and e-commerce (25). A recent discussion by Trincado-Munoz et al. (26) has noted that the professional services of AI have been expanded to the fields of justice and health, and that this expansion faces many challenges, including the explainability of AI technology, privacy, and human agency issues, which will disrupt the provision of AI services and have a negative impact on organizations, professionals, users and society (e.g., reduced access to services, marginalization and discrimination, malicious use of AI, inaccurate services and low service quality, data security vulnerabilities). Similarly, Aravazhi et al. (27) also argue that data privacy, algorithmic bias, and ethical dilemmas would cause AI users to worry about security and privacy.

Nevertheless, behind the use of intelligent agents as alternative service providers, security issues need to be inspected extensively. Recent research concerning AI security has focused on detecting malicious activities beneath the automated process, such as the flagrant or underhanded violation of privacy protection protocol, which gives rise to corresponding privacy concerns on the users’ part (26,27). However, the activation of such concerns does not merely hinge on the detected malice or violation of privacy (or any other basic rights) in a post hoc manner; instead, one’s privacy risk perceptions could proactively influence the following AI use experience. In the context of AI-generated news consumption, for example, privacy risk and bias perceptions of AI were found in negative relationships with attitude-consistent news exposure, through a lowered algorithm-accepting behavior; and bias perception showed a positive relationship with news-seeking behavior, while privacy risk perception was negatively linked to news-seeking behavior (28). If viewed broadly, such findings present meaningful insights into the necessity of enhancing users’ certain perceptions of AIGCs, beyond news outlets, as part of future citizens’ AI literacy, buttressed by corresponding interventions to foster a more personalized and trustworthy information environment on the user’s end. In this context, HMC researchers could investigate, for starters, how the prompt-based delivery of AIGCs—broadly defined here as the combination of algorithm authorship and recommendation—affects user perceptions of the content, especially when some expectations based on individual concerns (presumably as a function of their literacy in the aspect of interest) are violated.

In the context of smart healthcare, users’ privacy concerns could mostly center on their electronic health records that involve personal biomedical information or the like, accompanied by many ethical issues (29). It is likely that receivers’ initial skepticism or distrust about the risk of AI leaking their privacy leads them to repudiate the use of an AI agent for healthcare at the very beginning. On the other hand, it can also be tempting to assume that the pre-condition of acceding to the use of smart healthcare is the acceptance of privacy risks, or stated differently, untrustworthiness in privacy protection, in front of the potential benefits and convenience, as a trade-off choice. In some cases of using health apps, users could interact with a smart healthcare agent incognito. Therefore, by clarifying the operational principles of personal information protection, users’ privacy concerns can hopefully be minimized. Yet, users’ privacy concerns could play a considerable part in their perceptions of AI trustworthiness under other real-life circumstances. Thus, privacy concerns should be included in the later theoretical synthesis.

Besides, interacting with a chatbot to seek emotional support (namely, the relational usage, for the most part) has become a representative and integral part of AI usage, which directly points to the utility of AI in terms of, for instance, psychological treatments, and even presents a bold possibility of replacing therapist-client communication with a Human-AI “upgrade”. A body of research has examined the effects of emotional support on the user’s perceptions in the condition of communicating with an AI agent. For example, focusing on the reduction of stress and worry as the perceptual response, Meng and Dai (30) examined the role of human vs. AI as a chat partner providing emotional support in supportive communication. Meanwhile, even not for relational usage or supportive communication, the effects of emotional support may also matter in the form of caring, which is believed to influence the degree of a communicator’s trustworthiness (19).

In the medical context, “You can probably think of a doctor who knows her stuff and is honest but seems preoccupied or uninterested in you when you complain about medical problems. The physician undoubtedly gets low marks on your credibility scale, and her advice probably has little impact on you.” (18). Compared with a merciful expert giving advice full of compassion, a callous expert, showing a cruel disregard for the advice seeker, could be considered uncaring and thus not worthy of trust. In a real-life situation, however, expecting a caring physician could turn out to be a moot point, since the possession of medical expertise has made this profession somewhat irreplaceable—a patient will have to trust the doctor, whether kindness is shown or not, in the face of severe medical problems. However, if the same medical advice is conveyed by an AI agent, will the caring level of such messages influence the perceived trustworthiness and, in turn, the credibility? Bearing this question in mind, researchers may naturally seek to test the caring effects of human vs. AI in a medical context, perhaps manipulating the caring level or the agent type.

Focusing on AI agents in smart healthcare, the current discussion merely aims for a comparison between two different types of AI to level playing field for AI generating medical information upon human users’ request, without juxtaposing human experts beside AI. For one reason, the established authority of human doctors, as a social construct, could make it less of a fair game for AI counterparts, at least for now. To summarize these concerns, we can land on a primary conclusion that user trust in AI for medical information is not only contingent on perceived credibility, but also on the integration context, professional role, institutional constraints, and so forth. Yet, in this paper, the scope of interest has been narrowed down to one-to-one human-AI communication, mainly drawing on evidence regarding individual engagement with AI-generated health messages, especially the varied levels of acceptance and uptakes, predicted by AI features and personal factors on the users’ part.

Information exchange in AI-aided decision-making: AI literacy and transparency

To germinate specific studies concerning smart healthcare users, scholarly attention is needed for the following questions: What are the central issues in smart healthcare AI user research? How can researchers address the methodological challenges brought by the black-box nature of decision-making through machine learning and AI-powered automation? Probing into such questions, an integrative theory can be helpful, especially when constructed from a user-centered perspective. A choice mindset, as previously noted, on smart healthcare agents will be proposed, in turn guiding the propositions for future research and hopefully encouraging more discussions on this topic.

When going through a decision-making process without AI tools, one may experience heuristics, biases, and fallacies that may influence the decision-making. For the same task, an AI agent could step in and play a part in providing the options for a user to consider, with information around each option, or even form the final decision-making as a user’s choice. For making an AI-aided choice, rather than a choice made by AI on behalf of a user, examined by Endacott et al. (31), at least two types of AI agents should be helpful: expert system and recommender system. While the combination of these two systems can be easily found in practice and research (32), one approach to understanding their differences on the users’ part may be a dichotomy of their generated contents. An expert system can provide professional solutions, usually in the form of an article tailored for the user, as the product of algorithm authorship designed to reproduce the judgment of a human expert—an in-depth discussion can be found in Madni’s study (33). On the other hand, a recommender system mainly serves as an information filter to present personalized recommendations, or a set of options for the user’s reference, which can be understood as the function of algorithm recommendation (34). Their output can be considered both as advice in a broad meaning, while requiring different levels of cognitive resources on the users’ part and influencing the decision-making differently.

Given this output typology, users facing different decision-making tasks may choose to interact with different AI agents that are expected to best help them make the choice. However, an individual might not be very aware of an AI agent being an expert system or recommender system, until the acquisition of certain literacy. One’s prior knowledge about AI, or AI literacy, could inform their choice of AI agent and influence the interaction (35). It is also important to note the growing awareness to AI transparency in a variety of contexts. For example, while featured by algorithm-powered recommendations, social media platform’s lack of control over their content leads to a fundamental problem, i.e., misinformation. Facing misinformation, as users struggle with determining the truth, social media platforms should strive to empower users by making the algorithm recommendation more transparent and explainable (36). In such cases, the transparency of AI, rather than self-claimed accuracy, is likely to largely inform the credibility evaluation on the source. In general, the boundaries of AI research are being pushed forward, making the transparency issues increasingly complicated. Among other advances, AI for data mining is flourishing, supposedly able to enrich, in a revolutionary style, people’s means of acquiring knowledge. In the meantime, knowledge representation is believed to play an enormous part in agent-based computing, self-aware computing and so forth, implying the increasingly strengthened bonds between cognitive science and AI research (37). Put differently, it is natural for a user to wonder how the requested knowledge is learned and presented by an AI agent, as well as how the required expertise can be guaranteed, rather than simply accept the AI agent as a black box magically producing knowledge out of nowhere. In this sense, whether the content production process is explainable matters. And this leads the discussion back to the output dichotomy. Unlike the classic contrast of human vs. AI authorship (usually depicted as diametrically opposed), pre-programmed recommendation and algorithm authorship require more nuanced understandings and explanations to address their transparency. In the medical context, the recent emergence of accurate AI models for disease diagnosis further raises the possibility of AI-based clinical decision support (38), which represents an AI expertise type in terms of diagnosis (initial judgment) and prescription (authorship, yet without authorization for medication or treatment); meanwhile, another AI expertise type, mainly in charge of recommending the suitable health products to the right customers, has been widely used across situations, driven by the need for diversified health consumptions.

Conceptual modeling

One key objective of the current paper is to map a choice mindset on smart healthcare agents. From the viewpoint of behavioral economics, an important part of decision-making is the principle of opportunity cost, which allows a decision maker to compare the lost benefits and cost of making a certain decision with the gained benefits (39). Resonating with this principle, the ELM posits that people tend to spend more time and efforts (as the cost) processing messages of great personal relevance for securing potential benefits. Few would deny that health is a topic of great relevance to everyone. Yet, that does not necessarily mean that every AI-generated health message will be systematically processed in the context of smart healthcare. Also considering the emergence of decision-making as a key catalyst for advice adoption, at least two scenarios of smart healthcare usage can be considered, together as part of a possible categorical moderator. They are described as follows:

Scenario 1

Health product seeking task, in search of health product recommendation. When people aim to boost their state of health without an emergency, they may interact with a smart healthcare agent in a way that resembles online shopping. Accordingly, their expectation of the AI-generated information could be a personalized recommendation on several products for their choice. Overall, in such non-urgent, health-promotion contexts (e.g., seeking daily supplements or wellness products), users are guided by a promotion focus and value autonomy. They may expect the AI agent to act as a recommender, presenting a range of options rather than prescribing one specific choice.

Scenario 2

Medical solution seeking task, in search of medicine prescription or treatment plan based on primary diagnosis. When people feel slightly indisposed and need medical attention to a certain extent (not so much for requiring the immediate presence of human medical experts), they may choose to ask for personalized medical solutions from an AI agent, especially when no human agent is available. Therefore, solution-like information would be needed. Generally, in such moderately urgent contexts (e.g., addressing mild but concerning symptoms), users operate under a recovery focus and seek certainty. They may expect the AI agent to behave more prescriptively, offering direct guidance or solutions.

These two imagined scenarios, characterized by dramatically different levels of emergency, foreground the variations in user task types, which reflects a broad motivational difference between a health promotion focus and a health recovery focus. The expected role of the AI agent is likely to be recommender system (in Scenario 1) vs. expert system (in Scenario 2). In Scenario 1, when the AI agent supports users’ decision-making freedom through recommendation framing, perceived credibility could be enhanced; conversely, overly prescriptive responses in such low-risk contexts may trigger expectation violations and diminish credibility. In Scenario 2, when the AI agent provides authoritative, solution-driven advice aligned with users’ recovery goals, credibility may increase; a lack of clear direction or solution, however, may undermine the AI’s perceived competence and trustworthiness.

As maintained by Lee (40), instead of the mindlessness account, machine heuristics may function as a conditional moderator of source effect, mainly based on expectancy violation vs. confirmation regarding the contrast of human and machine. Similarly, a choice mindset on smart healthcare agents can be mapped. It comes as no surprise that not only for health decisions, in most situations, decision makers are far from being fully informed, infinitely sensitive to information, or completely rational. Therefore, decision-making strategies based on heuristics such as satisficing, representativeness, availability, framing, and anchoring, could be adopted. In this context, various information processing strategies for decision-making have been widely examined. Yet, the proposed concept map, as shown in Figure 1, only outlines several important points and routes, in pursuit of a theoretical integration and generalization.

Figure 1 A choice mindset map on smart healthcare agents. This model illustrates how users’ health literacy and AI literacy inform their perception of a medical emergency and uptake of smart healthcare agents, leading to different expectations: recommendations in relatively low-risk, promotion-focused conditions vs. prescriptions in relatively high-risk, recovery-focused conditions. Matches between AI deliveries and user expectations confirm the source fit and lead to source evaluation, while mismatches trigger expectation violation and re-confirmation seeking. This process highlights human-AI communication dynamics based on risk perception and motivational orientation. AI, artificial intelligence.

Health literacy and AI literacy, as a user’s prior knowledge, presumably determine the technology acceptance of a smart healthcare agent in a joint manner, by evaluating the medical emergency and the necessity for smart healthcare, as well as the potential costs of the usage at the pre-use stage. This paper, however, focuses on the credibility evaluation at the message acceptance level. Upon deciding to use a smart healthcare agent, a task-related expectation would be formed mainly based on the medical emergency, which can be divided into (I) expectation for recommendation in the situation of relatively low risk and (II) expectation for prescription in the situation of relatively high risk. While both situations can be considered personally relevant (which means, as postulated by the ELM, users would like to spend some time and efforts processing the messages in both situations), a focal difference should be noted, as implied in the descriptions of the two imagined scenarios: the former seeks to promote the health, with no perceived illness; yet, the latter seeks to get recovered, asking for medicines or solutions to illness.

It is assumed that only when the expectation is confirmed, or there is a scenario-expertise match, the source evaluation, including credibility evaluation, will be initiated; otherwise, users may want to keep looking for the task-relevant expertise until the match is confirmed—by rephrasing their prompts to the same agent (e.g., “maybe the AI misunderstood what I was asking about”) or just switching the agent (e.g., “maybe I got the wrong AI for the task”). For one possible reason, it can be easily understood that anyone who only asks for a simple recommendation would not like to waste time reading a long prescription that seems to exaggerate the health issues; on the other hand, prescription seekers would not be satisfied by simple recommendations on some medicines (such tasks can be easily done with a search engine), feeling that the AI agent was not taking their issues seriously. In this sense, a scenario-expertise mismatch is likely to reduce the source credibility to a degree that the user might think of changing the source. To examine this, several propositions will be developed in the next section.

Propositions for future research

Different facets of credibility will be influential in different social situations. This conclusion emerged from early studies of credibility, which found that results varied, depending on the ways in which researchers probed credibility and the circumstances in which the messages were delivered (18). Therefore, the preceding discussion has established two specific situations for examining the credibility of smart healthcare agent. Research on source credibility consistently highlights that users evaluate information more favorably when the expertise of the source aligns with their expectations within a given context (41,42). In health communication, patients show greater trust in medical advice when it comes from perceived health experts rather than generalists (43). In human-AI communications, perceived expertise remains crucial (44), but the AI agent’s role as either a “recommender” or a “prescriber” introduces complexity. In low-risk scenarios, users expect advisory, recommendation-based input, while in high-risk medical emergencies, users prefer authoritative, prescriptive guidance. When the AI delivery of information matches this expectation (recommendation in low-risk scenarios; prescription in high-risk scenarios), perceived credibility may increase because the AI’s “expert role” aligns with the user’s situational needs and mental model. Conversely, mismatches (e.g., a strong prescription in a low-risk scenario or vague recommendations in emergencies) could violate users’ expectations, triggering skepticism and expectation violation effects (45,46). While the importance of matched expertise is well-established for human sources, little empirical work has validated this mechanism in human-AI communication for medical information, where most users must infer expertise based on message framing alone without obvious human cues like academic degrees or professional appearance. In this context, built on the possible scenario-expertise relationship, the first proposition goes:

Proposition 1: Credibility will be higher in the scenario-expertise matched condition than in the scenario-expertise mismatched condition.

Following the framework of expectancy violation theory and empirical works on the impacts of various heuristics in credibility judgments, one can anticipate the influence of declared AI expertise and perceived caring on credibility perceptions. Future research may seek to take one step further to uncover how AI expertise types and caring levels may influence the credibility. Recapitulating some core assumptions and questions in the preceding discussion, what follows is another proposition along with two research questions for follow-up investigations.

A common expectation is that the computer-generated content should be bias-free (17). Thus, the declaration of algorithm authorship is likely to reinforce the expectations of objectivity from machines; yet, algorithm recommendation is often considered somewhat rigged, perhaps for the promotion of certain products. Is that so in the context of smart healthcare? Intriguingly, for news consumption, evidence suggests that the convenience perception of algorithm had a positive relationship with recommendation-accepting behavior, trust, and content-seeking behavior, respectively (28). If such findings could extend to the context of medicine recommendations or health suggestions, human doctors might succumb to algorithms, considering the relative inconvenience of accessing human services (without considering other issues such as professionalism, certified by medical licenses). However, if taking transparency into consideration, another question worth thinking is whether the trust in the human programmers behind an algorithm would influence a user’s choice of service in the same case. When it comes to governing machine learning for personalized medicine recommendation, the programming expertise of programmers (the foundation of a recommender system) vs. the medical expertise of doctors (the foundation of an expert system) could become a major concern, based on which the user’s choice would be made. To assuage a pain, for example, medicine can be prescribed from a doctor-simulating AI expert or recommended by a pre-programmed algorithm. For a recommendation algorithm to perform such tasks, a human programmer could first use natural language processing to analyze the user evaluation of existing drugs based on drug evaluation website data and thereby enable the algorithm to recommend the medicine, in line with the results of review analysis and the user’s condition. After explaining this mechanism/expertise to the user, will the user still trust a pre-programmed agent with his or her health? Or is it better if an expert system, capable of writing a prescription in a doctor-like style, is available? To address this, one research question can be set as a starting point:

Research question 1: Will credibility be higher in the prescription-receiving condition than in the recommendation-receiving condition? Or vice versa?

Moreover, as previously discussed, caring is theoretically important in determining trustworthiness and influencing credibility indirectly. In traditional interpersonal communication, caring cues such as empathy, attentiveness, and emotional warmth significantly boost perceived trustworthiness and subsequent credibility (47). Emerging studies suggest that these dynamics extend into human-AI interactions as well: AI agents perceived as empathetic and responsive evoke stronger emotional engagement and trust (48,49). Facing medical AI, users could be highly sensitive to the emotional quality of interactions, especially when dealing with health-related vulnerabilities. High-caring responses (e.g., empathetic affirmations, personalized follow-up suggestions) would likely enhance users’ perception of the AI as trustworthy. In turn, trustworthiness could act as a psychological bridge to overall credibility evaluations (50). By contrast, low-caring responses (e.g., curt answers, impersonal language) can trigger detachment, suspicion, and reduced acceptance, especially in emotionally charged health contexts where support and understanding are crucial. Although previous studies have explored AI empathy in mental health contexts (e.g., counseling bots), limited research directly examines caring in human-AI communication for medical information, let alone how trustworthiness may mediate the credibility outcome. Building off the existing studies on AI caring, we propose:

Proposition 2: Credibility will be higher in the condition of high-level caring than in the condition of low-level caring, which is mediated by trustworthiness.

When it comes to the possible interaction effects of expertise type and caring level on credibility, virtually no evidence can be found in previous literature to support any proposition. Therefore, another research question can be raised:

Research Question 2: How will the expertise type and the caring level interact to affect the credibility of the smart healthcare agent in different scenarios?

Discussion

While focusing on developing a conceptual model that extends the choice mindset into human-AI communication for medical information, we also encourage empirical work to further validate and extend our theoretical propositions. Considering the strength of experimental design in systematically manipulating AI expertise types to examine their impact on perceived credibility, the following discussion is a proposed research design, along with a critical reflection on the challenges and limitations that must be carefully addressed in future studies and practices.

Recommendations on research method, procedure, and tools

To test the propositions and address the research questions, a 2 (scenario: health product seeking task vs. medical solution seeking task) ×2 (declared expertise type: pre-programmed recommendation vs. algorithm-based prescription) ×2 (caring: caring vs. not caring) design can be adopted for an online experiment with participants drawn from a national sample and randomly assigned to one of the eight conditions that vary the task scenario, the expertise byline, and the caring of the health suggestion. Under each scenario, by manipulating the other two independent variables, four conditions can be created, as shown in Figure 2.

Figure 2 Recommended 2×2 manipulation table for each scenario. This figure presents a 2 (declared expertise type: pre-programmed recommendation vs. algorithm-based prescription) × 2 (caring: caring vs. not caring) between-subjects design in each scenario (health product seeking vs. medical solution seeking). Following this design, participants can be randomly assigned to one of eight conditions that vary scenario context, source framing, and perceived caring in the AI-delivered health suggestion. Under each scenario, four conditions are created by crossing expertise type and caring cues (Level 1, indicated by “X”, with no cues; Level 2, indicated by “O”, with cues). AI, artificial intelligence; PR, pre-programmed recommendation; AP, algorithm-based prescription.

The caring can be manipulated by two levels of caring phrases, varying in the use of emotion-laden expressions like “I feel sorry for you”, “I would be happy to hear from you again”, which could be pretested to ensure that the manipulation worked. The expertise type can be manipulated by two types of AI systems declared as a pre-programmed recommender system and an expert-like algorithm, along with the explanation on the expertise (i.e., how it works) respectively. Due to the possibility that some users with few transparency concerns may pay little attention to the expertise explanation, a manipulation check needs to be conducted. Only the participants who pass the manipulation check should be included in the subsequent analysis. As revealed in practices, while accurate AI for disease diagnosis could lower healthcare workloads, time and financial resources for gathering input data are usually limited (38), let alone security concerns regarding protected health information (PHI). Therefore, instead of asking participants to provide a great volume of real-life personal health data, which could be a daunting burden, a predetermined task could be assigned, in line with the scenario, while having nothing to do with the real-world health status of participants.

To cope with the inquiry task, every participant should be allowed to write a short prompt for the smart healthcare agent to ask for health advice (recommendation or prescription), as in a simulation of conducting the same task in the real world. After receiving and reading the advice, every participant then takes a survey for the measurement of variables of interest. For measuring the dependent variable, perceived credibility, the Source Credibility Scale developed by Ohanian (51), originally in the context of social marketing, could be adapted, consisting of two subscales for perceived expertise and trustworthiness respectively (the subscale for attractiveness can be removed). Notably, AI literacy and health literacy, as two possible covariates, could be gauged by asking participants to self-rate them on a Likert scale (since a systematic literacy test could be too challenging), along with common demographic variables.

Limitations and future directions

AI, since its marked proliferation in the mid-2010s, has significantly penetrated various sectors, becoming a transformative force that has necessitated the ongoing refinement of AI-based communication strategies and engaged a wide range of stakeholders. A general trend is, while AI research in terms of reasoning keeps increasing, perception-centered research is becoming more independent, chiefly contributing to the development of machine vision, agent-based computing, and cognitive robotics (52). However, it does not mean that reasoning and perception cannot be combined in research and practice. The connection between the two research areas should be an important focus of agent-based computing, cognitive robotics, integrated cognitive modeling, and so on, all falling under the umbrella of AI. Looking at such areas, Franklin (52) pointed out, knowledge-based expert systems were among the recent major AI achievements. Meanwhile, earlier achievements including recommender systems remain in service. The co-existence of these AI systems inspired us to conceive the present theory paper in the first place. On this topic, even though any full picture could be a fallacy for now, as with most theoretical works, the limitations of the current attempts are inevitable.

First, any generalization regarding the relationship between AI expertise type and credibility can be problematic. As clarified by O’Keefe (19), for low-credibility sources, they are not low in absolute terms, but relatively low in credibility. Stated differently, credibility is not an intrinsic source feature, but an outcome of subjective judgment. Thus, although the expected discussion by follow-up studies will be cast as a matter of how high- vs. low-credibility can be formed in the context of AI-based medical information seeking, the comparison will be limited to the contrast of an AI type relatively higher in credibility and a relatively lower one, not necessarily between two AI types that are absolutely high or low in credibility. That is, under no circumstances should one guarantee that recommender system will be considered more credible than expert system, or vice versa.

It is also natural to assume that a patient and his/her family member(s), if any, would engage in the communication with the smart healthcare agent for health advice, as it could be a critical decision to make together, since every family member is also a key stakeholder. However, the current modeling has merely considered individual subjects, thus failing to account for this close-to-life family scenario of smart healthcare use, let alone the dynamics within the group of “co-users”. Group decision-making, as discussed by some scholars (53), have been proven beneficial in certain cases, while the conversion from group decision-making into group thinking could give rise to a series of new issues worth academic attention. In this sense, future research should investigate the mechanisms of AI-aided group decision, especially the issues arising when the need for group decision activates group thinking among co-users during an AI-based medical information seeking process. Likewise, such human roles also need to be highlighted in a clinical case scenario. It is suggested that medical staff should be involved in the development and deployment of medical AI to help improve the acceptance of AI-generated medical information (54). For example, in emergency medicine, while AI-generated handoff notes can assist in care transitions, human-in-the-loop workflows are still necessary to manage risks with detailed and accurate LLM-generated notes (15). This aligns with our conceptual claim that user expectation calibration and human oversight are pivotal in achieving credible AI communication in high-stakes health contexts.

Attention to the user perceptions after the advice seeking process and the supported decision-making is also due. Upon a decision is made and an outcome ensues, hindsight bias, to name but one, may rear its ugly head, as people tend to believe that they could have been more predictable than they were on a past event (55), which may lead to negative post-decision results and thus warrants relevant research. Moreover, identifying a genuine role of expertise in determining AI credibility can be a challenging task, because isolating the effect of this source variation could be hard. Researchers must ensure that the expertise manipulation (especially when it is categorical, as proposed in this paper) is not confounded with variations in the contents generated by different expertise types. Given the confounding of the differential types of AIGCs, results may more plausibly be explained to reflect differences in the contents per se, like the recommended items vs. the prescribed solutions, which speaks to the necessity of more nuanced manipulations or mixed-methods approaches.

Concluding this conceptual paper with ethical concerns, it is noteworthy that the ethical and social risks associated with the deployment of LLMs in smart healthcare could use a deeper discussion. Drawing on a more comprehensive framework could help us highlight issues unaddressed in the current paper, such as epistemic uncertainty, the risk of bias perpetuation, and social manipulation, all of which are pertinent to medical AI governance and user protection, across various health communication contexts where trust and safety are paramount (56), thus requiring further notice. In this view, we have merely emphasized how some of these risks relate to our framework centered on perceived credibility, particularly around user trust calibration and source transparency, which can be a centerpiece for marketing-facing applications of medical AI, as trust management constitutes a top priority for responsive healthcare branding and personalized communication strategies (57). The current proprietary vs. open-source AI model divide, as can be noticed, is also rendering credibility issues even more complicated. While proprietary models (e.g., GPT-4) offer greater sophistication but operate in opaque ecosystems, which can hinder user transparency and accountability, open-source models (e.g., LLaMA, Mistral) may allow for more ethical auditing and alignment with public health values. We encourage future studies to investigate how such structural differences might affect the perceived authorship and credibility attribution in patient-facing applications, especially when recommendation origins are obfuscated by model design or deployment practices.

Conclusions

Dating back to the early stage of AI development, most AI systems were developed to help human cognitive processes by conducting certain cognitive functions for users. Currently, intelligent assistants have been integrated into our daily lives. It is becoming increasingly difficult to distinguish whether one’s decision-making is informed/influenced by computing technology or not. Particularly, the user experience of communicating with AI agents for medical information must be discussed with complexities in mind. Examining the relationship between computers/machines (as the information source) and human cognition (as the information processor) is a multifaceted task, as there are multiple factors coming into play. For a starter, this paper sheds light on the declaration and explanation of two AI expertise types, in consideration of transparency and caring to investigate their influences on perceived credibility. To this end, a choice mindset on smart healthcare agents has been mapped as the product of a theoretical synthesis, which calls for follow-up attention and validation. Though with the limitations discussed above, drawn from this paper, implications can be further discussed for perhaps a collaborative mode in future smart healthcare, with the algorithm in charge of the straightforward or simple facts and the human experts specialized at the process further up the chain to focus on what warrants urgent medical attention.

Acknowledgments

The adaptation of choice mindset—originally proposed to understand how individuals interpret others’ behaviors—into a framework for examining individuals’ own choices regarding smart healthcare agents was inspired by discussions during the 2024 NCCU-SNU-UTokyo Joint Symposium on AI-Human Communication. We are grateful for the constructive feedback and insightful suggestions on earlier versions of this paper after an abstract presentation by the first author in a session of the symposium noted before.

Footnote

Peer Review File: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2025-86/prf

Funding: This article was supported by the BK21 FOUR Program of the Ministry of Education of the Republic of Korea and the National Research Foundation of Korea (NRF- 4120200613754) to Z.L.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2025-86/coif). Z.L. received support from the BK21 FOUR Program of the Ministry of Education of the Republic of Korea and the National Research Foundation of Korea (NRF- 4120200613754). The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Jayaraman PP, Forkan ARM, Morshed A, et al. Healthcare 4.0: A review of frontiers in digital health. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 2020;10:e1350.
2. Huang Z, George MM, Tan YR, et al. Are physicians ready for precision antibiotic prescribing? A qualitative analysis of the acceptance of artificial intelligence-enabled clinical decision support systems in India and Singapore. J Glob Antimicrob Resist 2023;35:76-85. [Crossref] [PubMed]
3. Hamedani Z, Moradi M, Kalroozi F, et al. Evaluation of acceptance, attitude, and knowledge towards artificial intelligence and its application from the point of view of physicians and nurses: A provincial survey study in Iran: A cross-sectional descriptive-analytical study. Health Sci Rep 2023;6:e1543. [Crossref] [PubMed]
4. Lei Z, Huang Y. Reapproaching human-chatbot interaction for early literacy development: A learner-centered review of literature and technology prior to ChatGPT 3.0. The International Journal of Technology Innovation, and Education 2025;3:7-39.
5. Savani K, Markus HR, Naidu NV, et al. What counts as a choice? U.S. Americans are more likely than Indians to construe actions as choices. Psychol Sci 2010;21:391-8. [Crossref] [PubMed]
6. Bach TA, Khan A, Hallock H, et al. A systematic literature review of user trust in AI-enabled systems: An HCI perspective. International Journal of Human–Computer Interaction 2024;40:1251-66.
7. Laumer S, Maier C, Gubler FT. Chatbot acceptance in healthcare: Explaining user adoption of conversational agents for disease diagnosis. In: Proceedings of the 27th European Conference on Information Systems (ECIS), Stockholm & Uppsala, Sweden, June 8-14, 2019.
8. Ricci F, Rokach L, Shapira B. Introduction to recommender systems handbook. In: Ricci F, Rokach L, Shapira B, editors. Recommender systems handbook Boston, MA: Springer; 2011:1-35.
9. Ma A, Yang Y, Savani K. “Take it or leave it!” A choice mindset leads to greater persistence and better outcomes in negotiations. Organizational Behavior and Human Decision Processes 2019;153:1-12.
10. Savani K, Stephens NM, Markus HR. Choice as an engine of analytic thought. J Exp Psychol Gen 2017;146:1234-46. [Crossref] [PubMed]
11. Lee JER, Nass CI. Trust in computers: The computers-are-social-actors (CASA) paradigm and trustworthiness perception in human-computer communication. In: Trust and technology in a ubiquitous modern environment: Theoretical and methodological perspectives. IGI Global; 2010:1-15.
12. Petty RE, Briñol P. The elaboration likelihood model. In: Van Lange PAM, Kruglanski AW, Higgins ET, editors. Handbook of theories of social psychology. Sage; 2012:224-45.
13. Can E, Uller W, Vogt K, et al. Large Language Models for Simplified Interventional Radiology Reports: A Comparative Analysis. Acad Radiol 2025;32:888-98. [Crossref] [PubMed]
14. You S, Yang CL, Li X. Algorithmic versus human advice: Does presenting prediction performance matter for algorithm appreciation? Journal of Management Information Systems 2022;39:336-65.
15. Landman AB, Tilak SS, Walker GA. Artificial Intelligence-Generated Emergency Department Summaries and Hospital Handoffs. JAMA Netw Open 2024;7:e2448729. [Crossref] [PubMed]
16. Sundar SS. The MAIN model: A heuristic approach to understanding technology effects on credibility. In: Metzger MJ, Flanagin AJ, editors. Digital Media, Youth, and Credibility. Cambridge, MA: MIT Press; 2008;73-100.
17. Tandoc EC Jr, Yao LJ, Wu S. Man vs. machine? The impact of algorithm authorship on news credibility. Digital Journalism 2020;8:548-62.
18. Perloff RM. The dynamics of persuasion: Communication and attitudes in the 21st century. 8th ed. Routledge; 2023.
19. O’Keefe DJ. Persuasion: Theory and research. 3rd ed. Sage; 2016.
20. Schwenke N, Söbke H, Kraft E. Potentials and Challenges of Chatbot-Supported Thesis Writing: An Autoethnography. Trends High Educ 2023;2:611-35.
21. Wien AH, Peluso AM. Influence of human versus AI recommenders: The roles of product type and cognitive processes. Journal of Business Research 2021;137:13-27.
22. Lee EJ. What triggers social responses to flattering computers? Experimental tests of anthropomorphism and mindlessness explanations. Commun Res 2010;37:191-214.
23. Dehnert M, Mongeau PA. Persuasion in the age of artificial intelligence (AI): Theories and complications of AI-based persuasion. Human Commun Res 2022;48:386-403.
24. Franklin S. Artificial Intelligence, General. In: Runehov ALC, Oviedo L, editors. Encyclopedia of Sciences and Religions. Dordrecht: Springer; 2013:139-43.
25. Abdelhamid S, Mallari T, Aly M. Cybersecurity Awareness, Education, and Workplace Training Using Socially Enabled Intelligent Chatbots. In: The Learning Ideas Conference. Cham: Springer Nature Switzerland; 2023:3-16.
26. Trincado-Munoz FJ, Cordasco C, Vorley T. The dark side of AI in professional services. The Service Industries Journal 2025;45:455-74.
27. Aravazhi PS, Gunasekaran P, Benjamin NZY, et al. The integration of artificial intelligence into clinical medicine: Trends, challenges, and future directions. Dis Mon 2025;71:101882. [Crossref] [PubMed]
28. Lee S, Kang S. User understanding and perceptions of news recommendation algorithms: Relationships with attitude-consistent news exposure, news trust, and news-seeking behavior. Korean Journal of Journalism & Communication Studies 2024;68:348-85.
29. Keshta I, Odeh A. Security and privacy of electronic health records: Concerns and challenges. Egyptian Informatics Journal 2021;22:177-83.
30. Meng J, Dai Y. Emotional support from AI chatbots: Should a supportive partner self-disclose or not? Journal of Computer-Mediated Communication 2021;26:207-22.
31. Endacott CG, Leonardi PM. Artificial intelligence and impression management- Consequences of autonomous conversational agents communicating on one’s behalf. Human Communication Research 2022;48:462-90.
32. Walek B, Fojtik V. A hybrid recommender system for recommending relevant movies using an expert system. Expert Systems with Applications 2020;158:113452.
33. Madni A. The role of human factors in expert systems design and acceptance. Human Factors 1988;30:395-414.
34. Deshpande M, Karypis G. Item-based top-n recommendation algorithms. ACM Transactions on Information Systems 2004;22:143-77. (TOIS).
35. Long D, Magerko B. What is AI literacy? Competencies and design considerations. In Proceedings of the 2020 CHI conference on human factors in computing systems 2020:1-16.
36. Geels J, Graßl P, Schraffenberger H, et al. Virtual lab coats: The effects of verified source information on social media post credibility. PLoS One 2024;19:e0302323. [Crossref] [PubMed]
37. Naveed Uddin M. Cognitive science and artificial intelligence: simulating the human mind and its complexity. Cognitive Computation and Systems 2019;1:113-6.
38. Erion G, Janizek JD, Hudelson C, et al. A cost-aware framework for the development of AI models for healthcare applications. Nat Biomed Eng 2022;6:1384-98. [Crossref] [PubMed]
39. Hoskin RE. Opportunity cost and behavior. Journal of Accounting Research 1983;21:78-95.
40. Lee EJ. Minding the source: toward an integrative theory of human–machine communication. Human Communication Research 2024;50:184-93.
41. Hovland CI, Weiss W. The influence of source credibility on communication effectiveness. Public Opinion Quarterly 1951;15:635-50.
42. Pornpitakpan C. The persuasiveness of source credibility: A critical review of five decades’ evidence. Journal of Applied Social Psychology 2004;34:243-81.
43. Freimuth VS, Stein JA, Kean TJ. Searching for health information: The Cancer Information Service model. Public Health Reports 1989;
44. Glikson E, Woolley AW. Human trust in artificial intelligence: Review of empirical research. Academy of Management Annals 2020;14:627-60.
45. Burgoon JK. Interpersonal expectations, expectancy violations, and emotional communication. Journal of Language and Social Psychology 1993;12:30-48.
46. Burgoon JK, Hale JL. Nonverbal expectancy violations: Model elaboration and application to immediacy behaviors. Communication Monographs 1988;55:58-79.
47. Mayer RC, Davis JH, Schoorman FD. An integrative model of organizational trust. Academy of Management Review 1995;20:709-34.
48. Verhagen T, Van Nes J, Feldberg F, et al. Virtual customer service agents: Using social presence and personalization to shape online service encounters. Journal of Computer-Mediated Communication 2014;19:529-45.
49. Sundar SS. Rise of machine agency: A framework for studying the psychology of human–AI interaction (HAII). Journal of Computer-Mediated Communication 2020;25:74-88.
50. Fogg BJ, Tseng H. The elements of computer credibility. Proceedings of the SIGCHI conference on Human Factors in Computing Systems. 1999. doi: 10.1145/302979.30300.
51. Ohanian R. Construction and validation of a scale to measure celebrity endorsers' perceived expertise, trustworthiness, and attractiveness. Journal of Advertising 1990;19:39-52.
52. Franklin S. History, Motivations, and Core Themes. In: Frankish K, Ramsey WM, editors. The Cambridge Handbook of Artificial Intelligence. Cambridge, UK: Cambridge University Press; 2014:15-33.
53. Black D. On the rationale of group decision-making. Journal of Political Economy 1948;56:23-34.
54. Huo W, Yuan X, Li X, et al. Increasing acceptance of medical AI: The role of medical staff participation in AI development. Int J Med Inform 2023;175:105073. [Crossref] [PubMed]
55. Roese NJ, Vohs KD. Hindsight Bias. Perspect Psychol Sci 2012;7:411-26. [Crossref] [PubMed]
56. Weidinger L, Mellor J, Rauh M, et al. Ethical and social risks of harm from language models. 2021. arXiv preprint arXiv:2112.04359.
57. Senyapar HN. Healthcare Branding and Reputation Management Strategies for Organizational Success. Technium Soc Sci J 2024;55:26-53.