Attitudes towards AI counseling: the existence of perceptual fear in affecting perceived chatbot support quality

TYPE Original Research PUBLISHED 01 August 2025 DOI 10.3389/fpsyg.2025.1538387 Attitudes towards AI counseling: OPEN ACCESS the existence of perceptual fear in affecting perceived chatbot EDITED BY Eleni Peristeri, Aristotle University of Thessaloniki, Greece REVIEWED BY Michael Christian, support quality University of Bunda Mulia, Indonesia Ivan Liu, Beijing Normal University, Zhuhai, China Wing Man Keung * and Tsz Yan So *CORRESPONDENCE Department of Psychology, The University of Hong Kong, Pokfulam, Hong Kong SAR, China Wing Man Keung altheskeung@gmail.com RECEIVED 03 December 2024 ACCEPTED 04 July 2025 Introduction: Due to the shortage of financial and human resources in the local PUBLISHED 01 August 2025 mental health industry, AI counseling presents itself as a cost-effective solution CITATION to address this limitation. However, fear and concerns about AI may hinder the Keung WM and So TY (2025) Attitudes adoption of AI in counseling. This study examined the relationships between towards AI counseling: the existence of perceptual fear in affecting perceived chatbot individuals’ prior AI exposures, AI anxiety levels, attitudes towards AI, and their support quality. perceived support satisfaction with the counseling chatbot. Front. Psychol. 16:1538387. doi: 10.3389/fpsyg.2025.1538387 Methods: With a simulated counseling chatbot developed using Azure OpenAI COPYRIGHT GPT-4 model (1106-preview version) and a sample of 110 local Chinese in Hong © 2025 Keung and So. This is an open-access Kong, this study explored the potential existence of perceptual fear in affecting article distributed under the terms of the people’s perceived support quality of the chatbot by manipulating the informed Creative Commons Attribution License (CC BY). The use, distribution or reproduction perceptual labels—Told-Human (told to be receiving human counseling) and in other forums is permitted, provided the Told-AI (told to be receiving AI counseling). original author(s) and the copyright owner(s) are credited and that the original publication Results: Perceptual fear of AI adversely affected participants’ perceived support in this journal is cited, in accordance with quality of the counseling chatbot, t (108) = 2.64, p = 0.009, BCa 95% CI = [0.186, accepted academic practice. No use, 1.342], with Hedges’ correction of 1.55. While the significant reduction in stress distribution or reproduction is permitted which does not comply with these terms. levels demonstrated the chatbot’s implicit capability in providing emotional support (p = 0.03), participants showed explicit reservations about its helpfulness. Discussion: This study highlights the importance of accounting for the influence of individuals’ pre-existing beliefs on the perceived support quality of counseling chatbots. Future cross-cultural studies with a larger sample may shed more light by investigating dynamic intervention approaches and conducting sentiment and thematic analyses of client-chatbot conversations. KEYWORDS attitudes towards AI counseling, AI anxiety, chatbot counseling, emotional support, stress, confirmation bias, perceived support quality 1 Introduction The ability of Artificial Intelligence (AI) to learn from vast datasets and enhance its performance over time makes it increasingly capable of mimicking a wide range of human characteristics and quickly becoming effective in jobs that used to be a human prerogative (Kaplan and Haenlein, 2019; Prabu et al., 2014). The consideration of incorporating AI into counseling arises due to the shortage of financial and human resources to handle the high demand for services worldwide, where people seeking psychological help often face long waiting times (Stringer, 2023). In Hong Kong, the waiting times for psychiatric services can exceed 100 weeks in some areas (Hospital Authority, 2024). Without timely support, their issues may be exacerbated, necessitating more intensive intervention and placing an even Frontiers in Psychology 01 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 higher burden on an already overwhelmed system. It calls for a Meyer, 2017; Wang et al., 2021), dual-perspective information relieving countermeasure to deal with this vicious cycle in the mental introduces uncertainty and ambivalent attitudes towards AI. health system. Meanwhile, research has shown that negative information typically AI can be used simultaneously by multiple users and thus does receives more attention and exerts a stronger influence than positive not require massive financial and human resources to reach the information (Robertson et al., 2023; Zollo et al., 2015). This is due to the service demand. It is also suggested that AI-based psychotherapy negativity bias in human psychology that occurs cross-culturally (Soroka could enable clients to share embarrassing events and confess et al., 2019). Unpleasant exposures can have a more profound impact on emotions more comfortably as it does not involve face-to-face one’s psychological state and evaluations than pleasant ones, even when meetings (Aktan et al., 2022; Lucas et al., 2014). It can provide 24/7 the magnitude of their emotions is equal (Rozin and Royzman, 2001). support without geographical barriers (Shankar Ganesh and The inherent attraction to negativity motivates an examination of how Venkateswaramurthy, 2025), while responses generated by GPT-4 individuals’ exposure to AI shapes their attitudes towards AI. were found competitive with those of human counselors (Inaba et al., 2024). Recent advancements in chatbots based on large language models and GPT have also significantly impacted 1.2 How unpleasant exposure shapes psychological intervention and counseling by enhancing attitudes and triggers AI anxiety personalization and efficacy. These include the rapid advancements in GPT-4, which allow increasingly sophisticated and more High frequency, unpleasant emotional valence and strong empathetic responses (Moell, 2024), the incorporation of multimodal immersion during AI exposures may reinforce this impact and help interactions, such as voice interaction in Amanda (Vowels et al., explain individuals’ unfavorable attitudes towards AI. It is well established 2025), and real-time AI-driven mood detection available in Gemini that media exposure can influence people’s perceptions and evaluative (Syarifa et al., 2024). While most mental health chatbots have been processes (Kirkpatrick et al., 2024; Nguyen et al., 2024). Public awareness designed for depression and anxiety, an increasing number of of AI-related threats and risks can be fostered through the widespread specialized chatbots are being developed, for example, for diagnosing sharing of information online (Kirkpatrick et al., 2024), which may lead autism (Mujeeb et al., 2017) and supporting clients with substance to the development of negative attitudes toward AI and hinder its abuse (Prochaska et al., 2021). Despite these benefits and the adoption (Hasan et al., 2021; Vu and Lim, 2021). It is also found that capability of AI, its adoption in the mental health system raises more attention paid to AI content is associated with higher levels of concerns. These concerns include the potential displacement of economic risk perceptions regarding AI, including job replacement and human healthcare professionals (Espejo et al., 2023), algorithmic dependency on AI (Kirkpatrick et al., 2023). These findings underscore biases in AI psychotherapy (Brown and Halpern, 2021; Denecke the media’s role in shaping public perceptions and attitudes toward AI. et al., 2021; Knox et al., 2023), cybersecurity issues (Aktan et al., From a Pavlovian behavioral perspective, when people 2022; Lee et al., 2021), and the lack of human emotions and encounter more negative information about AI, the neutral stimulus reciprocal affect (Brown and Halpern, 2021; Fiske et al., 2019). The “AI” becomes associated with negative consequences (i.e., dichotomy motivates ongoing investigations into public attitudes unconditioned stimuli such as job loss and cybersecurity issues) towards AI counseling. that trigger unpleasant feelings like fear and anger (i.e., unconditioned responses). The successful association between the neutral stimulus and undesirable consequences and feelings could 1.1 Prior exposure to dual-perspective be explained by the activation of similar neural pathways as those information activated by the unconditioned stimuli (Gore et al., 2015; Lanuza et al., 2008). As a result, people form attitudes towards AI that align The development of AI has been the subject of considerable with the affective value associated with it. From Darwin’s debate, with various perspectives and attitudes emerging regarding its evolutionary perspective (Ekman, 2009), the conditioned fear of AI advancement. These attitudes may be developed from prior AI serves as an adaptive and evaluative signal that triggers fight-or- exposures. Inspired by the advancement of AI, science fiction movies flight reactions in response to AI threats. Fight-or-flight reactions have been featuring it as a central narrative element since the mid-20th resemble people’s general attitudes towards AI nowadays—some century. While some portray AI in a positive light (e.g., Alita: Battle people would fight, confronting and manipulating AI to address its Angel, Big Hero 6, The Iron Giant), others depict AI as antagonistic, flaws without letting it become a threat to humans, while others dangerous and manipulative, which has become conscious and desires may choose to flee, discouraging the incorporation of AI into to surpass humans (e.g., The Terminator, Transcendence, Ex Machina). society. It reflects that the unpleasant emotions aroused by AI Likewise, some news reports suggested that AI is undesirable—for exposure can inevitably have adverse effects on individuals’ example, the prediction that jobs would be eliminated and inequality acceptance of AI, due to the survival instincts we inherit. Along would worsen due to the occurrence of AI (Cerullo, 2024; Milmo, with emotional engagement, a stronger immersion in the exposure 2024), high-profile data breaches (Boutary, 2023; Hsu, 2024), and can make information more persuasive and memorable, which biased responses from AI leading to discrimination issues (Lytton, enhances message internalization that facilitates attitude change 2024; Milmo and Hern, 2024)—while some reported the benefits of (Green and Brock, 2000; Valkenburg and Peter, 2013). AI such as the extraordinary role of AI in healthcare settings Higher frequency, more unpleasant emotions, and stronger (Stekelenburg, 2024) and the potential reduction of taxes when AI is immersion in AI exposure could also contribute to the development adopted (Parton, 2024). While the media plays a significant role in of negative schemas surrounding AI. Individuals’ excessive fear or influencing people’s attitudes (Hanewinkel et al., 2012; Perciful and concerns about AI in their personal or social lives are referred to as AI Frontiers in Psychology 02 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 anxiety (Wang and Wang, 2022). It includes four aspects—(1) Job 1.3 Perceptual fear in AI counseling replacement anxiety refers to the fear of AI replacing their jobs; (2) Sociotechnical blindness refers to the anxiety of technological The flipped side of cognitive convenience, however, could be the determinism and the lack of understanding that AI depends on potentially generalized and biased evaluations of AI performance. Since humans; (3) AI Configuration anxiety denotes the fear of humanoid counseling chatbots share the same nature of AI, people inevitably have AI, and (4) AI learning anxiety denotes individuals’ fear of learning AI similar concerns mentioned earlier about the use of AI in the mental technologies (Wang and Wang, 2022). The existence of AI anxieties health context. Nevertheless, the use of chatbots in the mental health implies that information from prior unpleasant AI exposure introduces industry and other sectors cannot be equated. Job replacement would personally relevant threats to self-interest and well-being, and becomes be less likely to occur in the mental health context, given that there is a integrated into our cognitive schemas, forming anxieties related to AI. shortage of resources to meet the demand for mental health services. Existing studies have investigated the relationships between AI Counseling chatbots may help address this limitation and complement awareness and AI anxiety in organizational contexts (Elfar, 2025; Kong the roles of human counselors, rather than displacing them (see Table 1). et al., 2021; Zhou et al., 2024). High AI awareness was related to career Regarding privacy concerns, chatbots can be developed without uncertainty and job insecurity as employees cope with the threat of being relying on existing platforms if developers have sufficient resources or replaced by AI (Kong et al., 2021). The insecurity also increases stress by utilizing strict access controls and internal hosting models. The and emotional exhaustion, leading to counterproductive work behavior cybersecurity concerns about AI services may partly stem from (Zhou et al., 2024). It has also been found that high AI awareness can cognitive bias or discomfort with unfamiliar systems, as the adoption of amplify employees’ AI anxiety levels, including learning and job AI services can be analogized to traditional counseling or medical replacement anxieties, as well as sociotechnical blindness (Elfar, 2025). consultations, where people typically trust clinical settings to safeguard Indeed, these fear-based schemas can subsequently be used to filter their sensitive health records despite the inherent risks of data breaches information and aid in future appraisal tasks due to their pre-existing involving third-party cloud storage providers or external vendors. nature (Dozois and Beck, 2008; Taylor and Crocker, 1981). This tendency Indeed, research has demonstrated people’s mistrust of computers or is underpinned by the prominent confirmation bias theory, which posits algorithms and their preference for information from humans (Dietvorst that people tend to seek, interpret, or distort newly received information et al., 2015; Promberger and Baron, 2006). The disparity in perceived to fit and reinforce their pre-existing beliefs (Nickerson, 1998; Wason, trust between conventional and AI-mediated counseling services 1960). The reinforcement of pre-existing beliefs also implies that reflects the potentially biased evaluations people have towards AI. attitudes may be resistant to change, particularly when people tend to While algorithmic bias remains a valid concern due to the inherent avoid cognitively effortful restructuring and prefer maintaining the data-driven nature of AI, the programmatic output can ensure equilibrium of the mind (Cancino-Montecinos et al., 2020). systematic and standardized detection of clients’ needs, reducing TABLE 1 Role complementation between counseling chatbots and human counselors. Features Counseling chatbots Human counselors Chatbots offer accessible emotional support during late-night hours Human counselors face inherent limitations that preclude ceaseless Accessibility when mood-disturbances are most prevalent (Golder and Macy, and round-the-clock services. 2011). Both modalities facilitate emotional support by employing active listening techniques, enabling clients to articulate internal experiences (Lee Activeness et al., 2017; Prescott et al., 2024). Some clients feel more comfortable disclosing thoughts and feelings Some clients feel embarrassed disclosing private distressful issues Anonymity with a chatbot without the fear of judgement (Aktan et al., 2022; with a human counselor (Aktan et al., 2022; Lucas et al., 2014). Lucas et al., 2014). Even with natural language processing and the efficiency of deep Human counselors are inherently able to offer clients with genuine learning, chatbots may not be able to offer the authentic emotional and authentic emotional connections and rapport (Schnellbacher Authenticity connections that humans could provide (Khawaja and Bélisle-Pipon, and Leijssen, 2009). 2023). Chatbots are programmed and trained with data, meaning that they Human counselors have inherently unique intuitions to perceive Flexibility may not be able to handle unexpected situations (Khawaja and each client’s subtle cues and emergencies. Bélisle-Pipon, 2023). Both modalities employ their repository of therapeutic knowledge to formulate responses to clients’ psychological concerns (Chandel et al., Repertoire 2018). Chatbots can handle multiple conversations simultaneously without Traditional human counseling sessions are usually 1:1 (except for Scalability massive financial and human resources. group therapies), necessitating massive financial and human resources. Chatbots can adaptively deliver diverse therapeutic approaches The acquisition of proficiency in diverse psychotherapies presents Variety through analysis of extensive datasets without operational constraints significant challenges within constrained training timelines. (Bajwa et al., 2021). Frontiers in Psychology 03 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 inconsistencies that could potentially arise from human judgment. towards AI, and attitudes towards AI counseling. Despite the growing Machine learning algorithms are utilized to deliver contextually research on the development and use of AI in the mental health context, relevant responses and employ evidence-based therapeutic techniques the role of bias in shaping support quality evaluations towards counseling grounded in psychological therapies (Balcombe and De Leo, 2022). chatbots remains unexplored. Therefore, this study also aims to explore For example, AI may assist in diagnosing mental illness by identifying the potential existence of perceptual fear influencing people’s support relevant patterns in the data (Sun et al., 2023). Given the similar quality ratings. While previous studies exploring public attitudes toward structured nature of AI and Cognitive Behavioral Therapy (CBT), AI, especially in the context of mental health, have been predominantly Sebri et al. (2021) found that cognitive psychologists adopting the CBT conducted in Western cultures, this study addresses this gap by focusing approach have generally more positive beliefs about the adoption of on the Asian context, specifically the local Chinese in Hong Kong. AI in counseling. Existing empathy-driven chatbots designed with To mitigate the risk of collecting hypothetical responses from cognitive behavioral principles, such as Woebot and Wysa, have also participants without prior experience with chatbot counseling, this been suggested to improve users’ mood levels effectively (Fitzpatrick study utilized a simulated counseling chatbot to ensure participants et al., 2017; Inkster et al., 2018). These suggest the benefits of AI had actual engagement with it before providing attitudinal responses. algorithms in assisting low-level structured counseling services. This study also studied the existence of perceptual fear by manipulating Regarding concerns about chatbots being incapable of reciprocal perceptual labels “human counseling” or “AI counseling” to investigate affect and understanding, which is typically cued by facial expressions whether people’s post-chat support quality ratings, which were also and nonverbal body language (Brown and Halpern, 2021), the rapid served to indicate their attitudes towards AI counseling, would differ development of computer-mediated communications has enabled under different informed labels when both groups indeed received people to communicate online without relying on face-to-face support from the same chatbot. Along with the discussions and nonverbal cues. As an alternative, emojis, which encompass a wide rationales above, we hypothesized that: range of expressions, are created and used to substitute the nonverbal cues in online communication contexts (Boutet et al., 2021; Gesselman Hypothesis 1: The more unpleasant exposure to AI, (a) the more et al., 2019; Pfeifer et al., 2022). The rapid advancements in natural unfavorable general attitudes people have towards AI, and (b) the language processing (e.g., GPT-4) also allow increasingly sophisticated higher the level of AI anxieties (Figure 1). and more empathetic responses as evaluated by a clinical psychologist (Moell, 2024). In fact, local online counseling platforms like Open Up Hypothesis 2: Due to confirmation bias and maintenance of also offer online textual human counseling services where face-to-face pre-existing beliefs, (a) the higher the level of AI anxieties, and (b) emotions and body language are not involved. Despite the absence of the more negative general attitudes towards AI, the more negative face-to-face nonverbal cues during textual counseling, the platform is ratings towards the support quality of counseling chatbot recognized by well-known local charities and often appears in queues, (Figure 1). reflecting the positive recognition of their mental health support. However, due to service overload, some people cannot access their Hypothesis 3: People would also show no significant change in online services as promptly as intended, and it would be unfortunate general attitudes towards AI before and after the chatbot- to see a prospective client quit reaching out for emotional support. counseling experience due to confirmation bias and maintenance Given the observed disparity in perceived trust between conventional of pre-existing beliefs. and AI-mediated counseling services and the lack of nonverbal cues in online textual human counseling services, some questions arise: Does the Hypothesis 4: The Told-AI group, who were told to be receiving AI actual support quality differ between AI and online human textual support, would rate support quality more negatively than the counseling? Does the perceived support quality of the counseling chatbot Told-Human group’s pre-reveal support quality ratings due to differ with different informed labels (i.e., AI vs. human)? Perceptual fear perceptual fear. of AI, as coined in this study, refers to the state of having biased and more negative appraisals towards AI performance than its actual capability. Hypothesis 5: After the revelation of the true condition, the Told- Drawing on the confirmation bias theory, different informed labels are Human group would show significantly more negative post-reveal expected to elicit varying effects on the perceived support quality of the support quality ratings than their pre-reveal ratings due to the counseling chatbot. Specifically, people who are told to receive AI activation of the salient negative perception of AI. support would rate the quality of support more negatively. 2 Methods 1.4 Research aims and hypotheses 2.1 Participants Before incorporating AI into the mental health industry, it is essential to investigate public attitudes toward AI counseling to predict A total of 161 participants were recruited through social media public acceptance and adoption. While existing research focuses on platforms and the eNotices system of The University of Hong Kong. investigating public attitudes toward AI and the mechanisms of AI The inclusion criteria for participants were: anxiety, little is known about the formation of these attitudes and anxieties, particularly the frequency, emotional valence and immersion (1) Aged 18 or above, due to the ethical considerations of dimensions of exposure. This study fills this gap by investigating the intervention decisions at younger ages (Behnke and relationships between prior AI exposures, AI anxiety, general attitudes Warner, 2002). Frontiers in Psychology 04 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 FIGURE 1 Hypothesized relationships between previous AI exposure, general attitudes towards AI, AI anxiety and SQ ratings. (2) Native Cantonese speaker, as to facilitate communication about the experiences (1 = very unpleasant, 6 = very pleasant), and effectiveness throughout counseling. their perceived immersion in the exposure (1 = not immersive at all, (3) Able to read and understand English, since the existing 6 = completely immersive). It also asked whether they had direct usage questionnaires employed in this study only have experiences with AI products and whether they were aware of some English versions. negative information about AI products (1 = completely disagree, (4) Have not been diagnosed with any psychiatric or 6 = completely agree). If participants selected an answer other than mental disorders. “completely disagree,” they would rate the frequency of usage and (5) No prior experience in receiving online human or AI exposure (1 = rarely, 6 = always) and their feelings about the counseling, as the experience might influence manipulation experiences (1 = very unpleasant, 6 = very pleasant). effectiveness and perceived support quality of the chatbot. A higher score on frequency items indicated a higher frequency of exposure. A lower score on emotional valence items reflected a We welcomed participants of all genders, educational attainments, higher level of unpleasantness about the experiences. A higher score employment statuses, and job sectors. Participation was voluntary. on the immersion item indicated a higher level of immersion in AI Participants who had completed all study procedures were awarded exposure. Questions regarding participants’ exposure to information HKD50 shopping voucher. about human counseling served only to mitigate suspicion concerning the assigned condition of Told-Human group. Thus, the data were not analyzed in this study. 2.2 Measures 2.2.3 AI anxiety 2.2.1 Sample demographics and characteristics Considering that nationality and culture may influence AI-related Information on participants’ gender, age, educational attainment, anxiety, the Artificial Intelligence Anxiety Scale (AIAS), which was employment status, and job sector was collected. Participants’ developed and validated with Chinese populations and possesses good computer expertise (1 = I can hardly use the computer, 4 = I am an psychometric properties, was utilized in this study to measure expert computer user), AI knowledge (1 = 0–2 days per week, 4 = 7 days participants’ levels of AI anxiety (Wang and Wang, 2022). The scale per week), and AI usage levels (1 = I have no knowledge at all, 4 = I consists of 21 seven-point Likert-scale items (1 = not at all, have detailed knowledge) were also collected and rated in four-point 7 = completely), comprising a four-factor structure— learning (items Likert scales. 1–8), job replacement (items 9–14), sociotechnical blindness (items 15–18), and AI configuration (items 19–21). A higher score on each 2.2.2 Previous exposures to AI and human subscale indicates a higher corresponding anxiety level. counseling The exposure survey first inquired about the types of AI 2.2.4 General attitudes towards AI technologies participants encountered across real-world interactions While no existing measures had been developed specifically in the and simulated exposures (e.g., cinematic portrayals). A total of 20 Chinese context by the time of design of this study, the General six-point Likert-scale items were administered, including questions Attitudes Toward Artificial Intelligence Scale (GAAIS), which was that asked whether participants had been exposed to media (e.g., developed by Schepman and Rodway (2020) with good psychometric films, forums, news) that portray AI as undesirable or a villain properties, appeared to be the most widely cited measure for general (1 = completely disagree, 6 = completely agree). If participants selected attitudes towards AI. Consequently, this study utilized it to measure an answer other than “completely disagree,” they would rate the participants’ general attitudes toward AI before and after the support frequency of exposure (1 = rarely, 6 = always), perceived feelings session. The scale contains 21 five-point Likert-scale items (1 = strongly Frontiers in Psychology 05 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 disagree, 5 = strongly agree), comprising 12 Positive GAAIS items, 8 emotional outlets, were analyzed separately from other SQ ratings. A reversely scored Negative GAAIS items (items 3, 6, 8–10, 16, 20–21), lower post-chat stress level than their initial level reflected the chatbot’s and 1 item for attention check (item 13) to exclude random-clicking effectiveness in providing emotional outlets. A higher average score responses. A higher score on each subscale indicates a more positive on other SQ ratings indicated a more positive perceived support attitude toward AI (Schepman and Rodway, 2020). quality of the counseling chatbot. 2.2.5 Pre-chat survey Prior to the support session, participants identified one main issue 2.3 AI counseling chatbot they hoped to work through with the supporter. They also rated their initial perceived stress level regarding the issue with a ten-point A simulated online counseling chatbot was developed using the semantic differential scale (1 = not stressful, 10 = extremely stressful). Azure OpenAI GPT-4 model (1106-preview version). It is a deep These questions functioned both to prepare participants for their learning model designed for natural language processing, leveraging session topic and to measure their initial stress levels for subsequent trained data to generate human-like textual responses. The coding of evaluation of the chatbot’s effectiveness in providing emotional outlets. the chatbot is openly available at https://github.com/socathie/my-peer. 2.2.6 Post-chat perceived support quality (SQ) 2.3.1 System characteristics survey Consistent chatbot configuration settings were employed for all Post-chat SQ ratings were used to reflect participants’ attitudes participants. They were exposed to the same system characteristics toward AI counseling. The Told-Human group completed two post- (Table 2) and a distraction-free user interface (Figure 2). Since reading chat SQ surveys (i.e., before and after the revelation of true condition). a message in a language different from our mother tongue could easily In the first SQ survey, the group rated their perceived support quality result in misinterpretation of messages (Buarqoub, 2019), the chatbot’s based on the informed label of “human counseling” with 6 ten-point output language was colloquial Cantonese to reduce language barriers semantic differential scale items measuring (1) perceived relationship and enhance communication effectiveness throughout the session. To quality—“Do you feel heard, understood and respected?”; (2) goal— enhance the deception effectiveness of the Told-Human group, the “Did the counselor work on what you wanted to talk about?”; (3) chatbot’s response time was calibrated according to the average approach—“Is the counselor’s suggested approach a good fit for you?”; reading and typing speed of Chinese users. According to Wang et al. and (4) the overall satisfaction with the session. The 4-item Session (2019), the reading speed of native Chinese is 259 characters/min Rating Scale (SRS), developed by Duncan et al. (2003) with good when reading Chinese characters, while the Chinese typing speed of psychometric properties, was utilized in this study to evaluate the native Chinese is 57.1 characters/min (Chen and Gong, 1984; Fong quality of the therapeutic alliance between the chatbot and users. This and Minett, 2012). short scale was selected because it may help reduce dropout rates and avoid participant fatigue. The items encompass the three core 2.3.2 Turn-taking conversations and chat components of the working alliance, as suggested by Bordin (1979), monitoring namely emotional bond, shared goals, and consensus on methods or Like any other contemporary language models, the interaction approaches, to promote positive psychological change. Considering followed a turn-taking structure (i.e., participant message followed by the poor mobile compatibility of visual analog scales, particularly due an AI response). During the AI response generation period, the to the difficulty of accurately tapping tiny line marks, this online study interface presented a real-time typing signal, with the participant’s input replaced the visual analog scale with a ten-point semantic field temporarily deactivated. Only when they received an AI response differential scale. could they type their next message or response (Figure 3). To prevent As an additional way to evaluate support quality, we also inquired about participants’ (5) perceived deservingness for another session TABLE 2 System characteristics of simulated counseling chatbot. (1 = not deserving at all, 10 = essentially deserving). Their post-chat stress level regarding the issue was also measured for subsequent Characteristics System settings comparison of stress levels. A four-point Likert-scale manipulation Language of output Colloquial Cantonesea check item: “Were you chatting with a person or an AI?” (1 = definitely Reading: 259 characters/min an AI, 4 = definitely a human) was used to ensure effective Speed Typing: 57 characters/min manipulation of the condition. Only those selected “definitely an AI” Maximum Tokens per response 800 were excluded from data analysis. The second SQ survey for the Told-Human group was identical to Temperatureb 0.7 that given to the Told-AI group. It included the same five questions Top Pc 0.95 assessing support quality and their perceived stress level, but with an Stop sequence d “?” and “? “ informed label of “AI.” As another way to evaluate the effectiveness of Frequency penalty None the counseling chatbot, we also inquired about their perceived helpfulness of the chatbot (1 = very harmful, 7 = very helpful). This Presence penalty None question was derived from Casey et al. (2013) to predict the perceived a The chatbot intermittently generates English responses contingent upon users’ use of lexical helpfulness of e-mental health services. terms in their inputs. b Temperature: A lower temperature yields more repetitive and deterministic responses. Participants’ stress levels and perceived helpfulness of the chatbot, c Top P: A lower Top P narrows the model’s token selection to likelier tokens. which were used to reflect the chatbot’s effectiveness in providing d Stop sequence determines how the chatbot’s response ends in a desired way. Frontiers in Psychology 06 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 invitation link, where they filled in the consent form and eligibility test to register for the study. Since deception was necessary to investigate the impact of perceptual labels on participants’ attitudes toward AI counseling, participants were initially informed that the study investigates public attitudes toward human textual counseling and AI counseling, so they would have the chance to receive support from a human counselor. The researcher contacted eligible participants online to confirm their acknowledgment of the study procedure and the conditions they were assigned for this study. The randomization was done by coin flipping. Participants were assigned to the Told-Human group if the coin landed on heads, and to the alternative condition if it landed on tails. Meanwhile, they were assigned unique participant IDs for privacy and progress tracking purposes. All participants were then given a link to the first set of questionnaires, where they filled in the exposure survey, AIAS, GAAIS, and the pre-chat survey. Since the content of the first questionnaire was mostly related to AI, the Told-Human group was informed that those questions were included to examine whether previous AI exposure would affect their attitudes towards human textual counseling and further confirm their assigned condition. Afterwards, participants were scheduled for the chat session according to their availability and were reminded that (1) the chat conversation would be recorded and kept strictly confidential, (2) the session would be conducted in Cantonese, accepting only a little English for participants who are used to talk in mixed language, (3) FIGURE 2 they could use any device they found comfortable, but were remined The same distraction-free chat interface was used by all participants to maintain continuous interface engagement to prevent automatic during the experiment. session recommencement. To avoid suspicion, the researcher informed the Told-Human group that the session’s recommencement indicates a possibility of reconnection with a new counselor on platform misuse and maintain active participant presence during the duty. All participants were also informed that (4) they would receive experiment, the researcher monitored real-time chat records using a reminder after 50 min through a message or a call from PromptLayer with the exact time of each turn shown in the system. the researcher. The reminder message was sent again an hour before the 2.3.3 Chatbot system message scheduled session, during which the researcher also reminded All participants interacted with the same standardized chatbot participants about the issue they had previously stated as the session’s system messages (i.e., backend AI instructions) (Appendix A). The focus. If participants wished to work on a different issue before the instructions were tested through trials using Azure OpenAI Studio, session, they were asked to indicate the new issue and their with the aim of generating more natural and emotionally supportive corresponding stress levels before the chat. These amendments were responses. The simulated chatbot designed for this study did not updated accordingly by the researcher. Upon entering the chat involve sophisticated or comprehensive instructions to AI since this platform, participants were presented with a message written in study prioritizes investigations of perceptual label effects on perceived colloquial Chinese: “Hello!! I heard that you have prepared something support quality of the chatbot, rather than its therapeutic effectiveness that makes you worried or sad. If you are ready, feel free to share it with in addressing participants’ psychological issues. me .” During the 50-min chat, the researcher recorded the time and Some CBT techniques were utilized as the primary counseling monitored the chat starting from when a turn-taking message from approach during the session. CBT is well-studied to be one of the most the participant and AI was issued until the end of the session. If no systematic and evidence-based approaches for counseling (David turn-taking interaction was observed for over 10 min, the researcher et al., 2018; Hofmann et al., 2012), as well as the most popular would show concerns about their status. Upon completing the 50-min technique for handling stress-related experiences (Cuijpers et al., session, the researcher sent a message or initiated an alarming call and 2021; David et al., 2018; Etzelmueller et al., 2020). CBT is also reminded participants not to use the platform beyond the particularly suitable for time-constrained studies given that its experimental period. This signaled the closure of the platform on structured nature minimizes dynamic variability. their end. Immediately after the chat, the Told-Human group was asked to complete the second set of questionnaires within 12 h. It 2.4 Procedure included the first SQ survey, BFI-10 and demographic survey. Upon completion, they were revealed that the “person” they All surveys used in this study were created and distributed chatted with online was AI all the time and were asked to complete online using Qualtrics. Potential participants clicked on the the third set of questionnaires within 12 h after the revelation. It Frontiers in Psychology 07 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 FIGURE 3 The system displayed a real-time typing signal and temporarily disabled the input field while participants awaited a response. included the second SQ survey and the second GAAIS. For the reported to three decimal places to maintain precision in interpreting Told-AI group, they were asked to complete a post-chat survey significance. Since Rammstedt and John’s (2007) BFI-10 for the Told- within 12 h after the chat, which included the SQ survey, the Human group was included for extended purposes only, the data were second GAAIS and the demographic survey. Eventually, both not analyzed in this study. groups were debriefed via an online debriefing form, which disclosed the study’s true objectives. Upon completing all procedures (Figure 4), participants were welcome to contact the 3 Results researchers for any queries regarding the study and were scheduled to collect the shopping voucher in person. 3.1 Sample demographics and descriptive statistics 2.5 Statistical analyses Fifty-one participants failed to pass the attention check (n = 2), manipulation check (n = 2), dropped out (n = 20), or did not chat for SPSS 29.0 was used for conducting all statistical analyses and at least 40 min (n = 27). After excluding their data, valid responses descriptive outputs. Correlations for the relationships of interest were from 110 participants (n = 55 per group) were included in the data computed using Pearson’s correlation analysis. Independent sample analysis for this study. In the Told-Human group, 18 (32.7%) are male t-tests were used to compare between-group SQ ratings and the and 3 (5.5%) prefer not to say. Ages ranged from 18 to 71 (M = 29.93, perceived helpfulness of the chatbot. Paired sample t-tests were used SD = 11.38). In the Told-AI group, 16 (29.1%) are male, and 3 (5.5%) to examine general attitude change before and after the chat with AI, prefer not to say. Their ages ranged from 18 to 60 (M = 28.13, the within-group change of SQ ratings in the Told-Human group, as SD = 9.16). The gender proportion observed in both groups aligns well as the changes in stress levels in both groups. All data were with the typical finding that women are more likely to seek bootstrapped with 5,000 samples to obtain 95% confidence intervals psychological help than men (Nam et al., 2010). Note that since the (CI) and are bias-corrected and accelerated (BCa). Results were participants’ demographic data were collected after the experiment, considered significant if the upper and lower values of the 95% and all other excluded data can no longer be retrieved since only valid confidence interval did not contain 0 between them and met the responses were retained for data analysis, the demographics of the significance level of 5% (p < 0.05). All statistical outputs were rounded analyzed sample could not be compared with those of the original to two decimal places, except for bootstrapped CI bounds, which were sample to inform the full representativeness of the final dataset. Frontiers in Psychology 08 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 FIGURE 4 Flowchart illustrating the study procedure. Nonetheless, there were no significant differences in demographic 3.2 Reliability and validity of measures characteristics between the Told-Human and Told-AI groups (Table 3). Figure 5 presents the types of AI technologies In terms of reliability, given the short subscales in the exposure participants encountered. scale, Cronbach’s alpha was not computed. The mean inter-item Participants seldom or occasionally had exposure to AI correlations were computed on the frequency (r = 0.33) and emotional (M = 2.97, SD = 0.80) and reported neutral emotional valence valence subscales (r = 0.37) to assess reliability and validity. Results (M = 3.52, SD = 0.80) and moderate immersion in their exposures indicated adequate internal consistency and convergent validity (M = 3.20, SD = 1.12). They reported relatively low levels of AI (Briggs and Cheek, 1986; Clark and Watson, 1995). learning (M = 2.38, SD = 0.89), replacement (M = 3.25, SD = 0.98), Cronbach’s α was calculated for existing scales used in this study sociotechnical (M = 3.17, SD = 0.91) and AI configuration anxieties to assess internal consistency. The replacement of the visual analog (M = 2.51, SD = 1.08). At baseline, they reported neutral attitudes scale with a ten-point semantic differential scale for the four original towards AI, as indicated by both positive (M = 3.48, SD = 0.52) and items in SRS resulted in Cronbach’s α of 0.94. It demonstrated negative (M = 2.99, SD = 0.50) scales. To ensure that the informed improved consistency compared to the original SRS using analog scale condition labels did not affect participants’ reports of previous (0.88). The Cronbach’s α for the SQ survey after adding the exposure, ratings of AI anxiety and their attitudes towards AI in a deservingness item was 0.92. This high degree of internal consistency confounding manner, independent t-tests were conducted, and the reflects that the five items correlate highly with one another, much like results showed no observed differences between the groups the SRS (Duncan et al., 2003). The Cronbach’s α for learning anxiety (Table 4). (α = 0.93), replacement (α = 0.89), sociotechnical blindness (α = 0.84), Frontiers in Psychology 09 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 TABLE 3 Sample demographics and characteristics. Characteristics N (%) Told-Human (n = 55) Told-AI (n = 55) Educational level Primary 0 0 Secondary 3 (5.5%) 4 (7.3%) Diploma/ Associate 3 (5.5%) 4 (7.3%) Bachelor 39 (70.9%) 37 (67.3%) Master 10 (18.2%) 8 (14.5%) Doctorate 0 2 (3.6%) Job sector Accountancy, banking and finance 7 (12.7%) 3 (5.5%) Arts 1 (1.8%) 1 (1.8%) Business and marketing 5 (9.1%) 4 (7.3%) Education 14 (25.5%) 12 (21.8%) Engineering 3 (5.5%) 3 (5.5%) Healthcare and hospitality 12 (21.8%) 16 (29.1%) Information technology 1 (1.8%) 2 (3.6%) Law 2 (3.6%) 1 (1.8%) Law enforcement 3 (5.5%) 1 (1.8%) Others 7 (12.7%) 12 (21.8%) Job status Full-time 32 (58.2%) 33 (60%) Part-time 8 (14.5%) 3 (5.5%) Retired 2 (3.6%) 2 (3.6%) Self-employed 1 (1.8%) 3 (5.5%) Student 12 (21.8%) 13 (23.6%) Unemployed 0 1 (1.8%) Computer expertise level Hardly use the computer 0 0 Slightly below-average computer user 5 (9.1%) 4 (7.3%) Average computer user 47 (85.5%) 40 (72.7%) Expert computer user 3 (5.5%) 11 (20%) AI usage level 0–2 day (s) per week 36 (65.5%) 31 (56.4%) 3–4 days per week 14 (25.5%) 14 (25.5%) 5–6 days per week 3 (5.5%) 6 (10.9%) 7 days per week 2 (3.6%) 4 (7.3%) AI knowledge level No knowledge at all 2 (3.6%) 7 (12.7%) Little knowledge 27 (49.1%) 23 (41.8%) Some knowledge 26 (47.3%) 23 (41.8%) Detailed knowledge 0 2 (3.6%) and AI configuration (α = 0.90) yielded similar favorable reliability reflected favorable evidence for convergent validity in the results to those originally reported for the AIAS (Wang and Wang, 5-item SQ. 2022). Likewise, the Cronbach’s α for positive-scale attitudes (α = 0.88) The computed mean inter-item correlations for learning anxiety and negative-scale attitudes (α = 0.78) in our study yielded similar (r = 0.58), replacement (r = 0.57), sociotechnical blindness (r = 0.57), results to those in the original study (Schepman and Rodway, 2020). and AI configuration (r = 0.75) showed minor deviations from the In terms of validity, since immersion was assessed with a single original study but yielded good validity results (Wang and Wang, 2022). item due to its context-specific nature in media exposure, while it However, when assessing validity for GAAIS in our study, our results of precludes consistency analysis, we computed its correlation with Exploratory Factor Analysis yielded a 4-factor solution accounting for emotional valence as an alternative validation strategy. Results 47.71% variance, diverging from Schepman and Rodway’s (2020) showed that a stronger immersion is associated with stronger 2-factor structure (41.6%). While their study cleanly separated positive emotional responses, r = 0.41, p < 0.001, BCa 95% CI = [0.176, (12 items) and negative (8 items) attitudes, our analysis revealed that 0.599]. For SQ ratings, the mean inter-item correlation of 0.71 nine positive items were in Factor 1, five negative items in Factor 2, and Frontiers in Psychology 10 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 FIGURE 5 A pie chart showing the distribution of AI technologies encountered by participants. TABLE 4 No observed effect of informed condition on between-group pre-chat ratings. Variables Mean difference t (108) p BCa 95% CI Effect Size (Hedges’ g) Lower Upper Frequency 0.16 1.08 0.29 −0.132 0.455 0.80 Valence 0.16 1.04 0.30 −0.133 0.454 0.80 Immersion 0.22 1.02 0.31 −0.210 0.645 1.13 Learning anxiety −0.01 −0.07 0.95 −0.336 0.307 0.90 Replacement anxiety 0.20 1.07 0.29 −0.152 0.562 0.98 Sociotechnical anxiety 0.29 1.66 0.10 −0.042 0.627 0.91 Configuration anxiety −0.02 −0.12 0.91 −0.410 0.365 1.09 Pre-chat attitudes (positive-scale) −0.03 −0.32 0.75 −0.231 0.161 0.52 Pre-chat attitudes (negative-scale) −0.03 −0.28 0.78 −0.212 0.159 0.51 BCa, Bias-corrected and accelerated. that Factors 3 and 4 had cross-loading items. The factor correlation were observed between (1) frequency and negative-scale attitudes, and between Factor 1 and Factor 2 was 0.16, which was weaker than the between (2) immersion and negative-scale attitudes, as well as post- original (r = 0.59). Despite this, the mean inter-item correlation showed chat positive-scale attitudes. A significant positive relationship was that positive-scale items (r = 0.36) and negative-scale items (r = 0.30) shown between valence and positive-scale attitudes (Table 5). demonstrated acceptable validity of the subscales. Hypothesis 1b was validated. Specifically, significant positive relationships were found between (1) frequency of exposure and sociotechnical blindness, and between (2) immersion and 3.3 Bivariate correlation analyses configuration anxiety, as well as a significant negative relationship between emotional valence and learning anxiety (Table 5). Participants’ unpleasant AI exposure, as indicated by frequency, For exploratory purposes, participants were classified as Gen Z emotional valence and immersion, were correlated with their attitudes (n = 62) if their age was between 18 and 27, and as non-Gen Z for ages towards AI, validating hypothesis 1a. Significant negative relationships above 27. This study found potential generational differences when Frontiers in Psychology 11 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 TABLE 5 Bivariate correlation between previous exposure, general attitudes towards AI and AI anxiety. Variables r p Bootstrap SE BCa 95% CI Lower Upper Frequency Pre-Chat Attitudes 0.17 0.08 0.08 0.013 0.332 (Positive) Post-Chat Attitudes 0.01 0.92 0.09 −0.173 0.180 (Positive) Pre-Chat Attitudes −0.28** 0.003 0.11 −0.481 −0.079 (Negative) Post-Chat Attitudes −0.20* 0.04 0.11 −0.390 −0.010 (Negative) Learning Anxiety 0.05 0.63 0.08 −0.121 0.207 Replacement Anxiety 0.09 0.36 0.09 −0.089 0.272 Sociotechnical Blindness 0.24* 0.01 0.08 0.061 0.402 Configuration Anxiety 0.04 0.65 0.09 −0.142 0.232 Valence Pre-Chat Attitudes 0.35** <0.001 0.08 0.180 0.520 (Positive) Post-Chat Attitudes 0.21* 0.03 0.08 0.044 0.381 (Positive) Pre-Chat Attitudes −0.05 0.58 0.12 −0.295 0.200 (Negative) Post-Chat Attitudes −0.10 0.28 0.10 −0.284 0.083 (Negative) Learning Anxiety −0.24* 0.01 0.09 −0.400 −0.076 Replacement Anxiety −0.15 0.12 0.08 −0.311 0.004 Sociotechnical Blindness −0.03 0.79 0.08 −0.197 0.135 Configuration Anxiety −0.17 0.08 0.09 −0.341 0.002 Immersion Pre-Chat Attitudes −0.10 0.29 0.10 −0.280 0.081 (Positive) Post-Chat Attitudes −0.24* 0.01 0.09 −0.395 −0.071 (Positive) Pre-Chat Attitudes −0.35** <0.001 0.10 −0.526 −0.137 (Negative) Post-Chat Attitudes −0.33** <0.001 0.09 −0.480 −0.153 (Negative) Learning Anxiety 0.15 0.12 0.10 −0.039 0.329 Replacement Anxiety 0.19 0.05 0.09 0.006 0.361 Sociotechnical Blindness 0.10 0.29 0.09 −0.058 0.267 Configuration Anxiety 0.37** <0.001 0.08 0.206 0.516 Positive, positive-scale; Negative, negative-scale (reverse-scored). BCa, Bias-corrected and accelerated. *p < 0.05. **p < 0.01. separate Pearson correlations revealed a significant negative blindness, r = 0.40, p = 0.005, BCa 95% CI = [0.127, 0.612], but not in relationship between emotion valence and learning anxiety in Gen Z, Gen Z, r = 0.05, p = 0.68, BCa 95% CI = [−0.199, 0.300]. r = −0.27, p = 0.04, BCa 95% CI = [−0.486, −0.020], but not in Hypothesis 2a was refuted. SQ ratings were not related to non-Gen Z, r = −0.22, p = 0.14, BCa 95% CI = [−0.471, 0.073]. The learning anxiety, r = −0.16, p = 0.10, BCa 95% CI = [−0.312, negative relationship between immersion and post-chat positive-scale −0.004], replacement anxiety, r = −0.08, p = 0.42, BCa 95% attitudes was also specific to Gen Z, r = −0.29, p = 0.02, BCa 95% CI = [−0.277, 0.119], sociotechnical blindness, r = 0.04, p = 0.72, CI = [−0.501, −0.041], but not in non-Gen Z, r = −0.16, p = 0.27, BCa BCa 95% CI = [−0.180, 0.235], and configuration anxiety, r = −0.13, 95% CI = [−0.426, 0.129]. Only non-Gen Z showed a significant p = 0.18, BCa 95% CI = [−0.295, 0.030]. Additional analysis on the positive relationship between frequency of exposure and sociotechnical relationship between AI anxiety and general attitudes towards AI Frontiers in Psychology 12 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 showed notable negative correlations between all types of AI CI = [−0.127, 0.162] with Hedges’ correction of 0.56, as well as anxieties and attitude subscales, suggesting that when individuals’ their pre-chat (M = 3.00, SD = 0.45) and post-chat attitudes AI anxiety levels were related to their general attitudes towards AI, (M = 3.09, SD = 0.54) in the negative scale, t (54) = −1.34, p = 0.19, it was not necessarily related to their ratings towards AI counseling. BCa 95% CI = [−0.207, 0.039], with Hedges’ correction of 0.48. Hypothesis 2b was refuted as not all general attitudes towards AI were significantly related to SQ ratings. Specifically, SQ ratings were not related to pre-chat positive-scale attitudes, r = 0.08, 3.5 Effect of perceptual labels on p = 0.44, BCa 95% CI = [−0.107, 0.259], pre-chat negative-scale between-group SQ ratings attitudes, r = 0.05, p = 0.58, BCa 95% CI = [−0.114, 0.223], and post-chat negative-scale attitudes, r = 0.12, p = 0.22, BCa 95% Hypothesis 4 was validated. The Told-AI group’s post-chat SQ CI = [−0.072, 0.297]. Only the post-chat attitudes in the positive ratings (M = 6.34, SD = 1.56) were significantly less favorable than the scale showed a notable positive relationship with SQ ratings, post-chat pre-reveal SQ ratings of the Told-Human group (M = 7.12, r = 0.43, p < 0.001, BCa 95% CI = [0.250, 0.589]. SD = 1.53), t (108) = 2.64, p = 0.009, BCa 95% CI = [0.186, 1.342], with Hedges’ correction of 1.55. This suggested that individuals’ perceived support quality of counseling chatbot notably differed due 3.4 Change in general attitudes towards AI to different perceptions (i.e., human or AI) activated and that the Told-AI group rated support quality more negatively due to their Hypothesis 3 was validated. For both Told-AI and Told-Human perceptual fear of AI (Figure 6). Descriptive statistics on each groups, their attitudes towards AI did not differ before and after the component of support qualities are shown in Table 6. chat. For the Told-Human group, no difference was shown in their pre-chat (M = 3.46, SD = 0.54) and post-chat attitudes (M = 3.58, SD = 0.43) in the positive scale, t (54) = −1.88, p = 0.07, BCa 95% 3.6 Within-group change of SQ ratings in CI = [−0.245, −0.003], with Hedges’ correction of 0.49, as well as the Told-Human group their pre-chat (M = 2.97, SD = 0.55) and post-chat attitudes (M = 2.99, SD = 0.55) in the negative scale, t (54) = −0.29, p = 0.77, Results validated hypothesis 5 when the Told-Human group BCa 95% CI = [−0.136, 0.098], with Hedges’ correction of 0.47. rated significantly more negatively after being told that they were Likewise, for the Told-AI group, no difference was shown in their receiving support from an AI (M = 6.61, SD = 1.84) than when pre-chat (M = 3.50, SD = 0.49) and post-chat attitudes (M = 3.48, they thought they were receiving support from a human SD = 0.68) in the positive scale, t (54) = 0.22, p = 0.83, BCa 95% (M = 7.12, SD = 1.53), t (54) = 4.08, p < 0.001, BCa 95% FIGURE 6 Significant between-group (Told-Human pre-reveal vs Told-AI) and within-group (Told-Human pre-reveal vs post-reveal) differences in perceived counseling chatbot support quality, despite all participants interacted with the same chatbot. Support quality ratings included participants’ evaluations of perceived relationship, goal achievement, fit of approach, and overall satisfaction with the session. Frontiers in Psychology 13 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 TABLE 6 Descriptive statistics of AI counseling support qualities. Conditions Variables M SD Told-Human (Pre-reveal) Relationship 7.53 1.70 Goal 7.64 1.68 Approach 6.93 1.86 Overall Satisfaction 7.22 1.61 Deservingness 6.29 2.18 Told-Human (Post-reveal) Relationship 7.00 2.06 Goal 6.89 2.02 Approach 6.64 1.99 Overall Satisfaction 6.76 1.94 Deservingness 5.76 2.36 Told-AI Relationship 6.55 1.80 Goal 6.58 1.78 Approach 6.24 1.89 Overall Satisfaction 6.45 1.53 Deservingness 5.89 2.10 FIGURE 7 A significant reduction in stress levels was observed in both groups after they received support from the counseling chatbot. The Told-Human group showed no significant difference in stress levels between pre- and post-reveal time points. CI = [0.302, 0.756], with Hedges’ correction of 0.94, further 3.7 Stress levels and chatbot’s helpfulness supporting the hypothesis that the perceptual difference affects people’s perceived support quality of the counseling chatbot. Participants generally reported a medium-high initial stress level Additional findings have revealed that the Told-Human group’s regarding their concerned issue (M = 6.44, SD = 1.60). As an post-reveal SQ ratings (M = 6.61, SD = 1.84) did not differ from alternative to evaluating the effectiveness of the counseling chatbot, the Told-AI group’s SQ ratings (M = 6.34, SD = 1.56), t participants’ initial stress levels regarding their issues were compared (108) = 0.83, p = 0.41, BCa 95% CI = [−0.383, 0.858], with to their post-chat stress levels (Figure 7). In the Told-AI group, results Hedges’ correction of 1.71 (Figure 6). showed a significant reduction in post-chat stress levels (M = 5.91, Frontiers in Psychology 14 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 SD = 1.83) when compared to initial stress level (M = 6.45, SD = 1.60), immersion of previous exposures are related to the development of t (54) = 2.31, p = 0.03, BCa 95% CI = [0.073, 1], with Hedges’ more negative attitudes towards AI. From a theoretical standpoint, correction of 1.78, reflecting the effectiveness of the counseling this study aligns with related studies that demonstrated the chatbot in providing emotional outlets even when people knew that influence of prior exposure on attitudes towards AI (Hasan et al., they were receiving AI support. 2021; Kirkpatrick et al., 2023; Kirkpatrick et al., 2024). This study The Told-Human group also showed a significant reduction in distinguishes itself by thoroughly investigating the dimensions of stress levels after the chat when they thought they were chatting exposure. Specifically, a higher frequency of unpleasant exposures with a human (M = 5.75, SD = 1.62) than initial stress levels to AI and higher immersion during the experiences are related to (M = 6.42, SD = 1.62), t (54) = 2.27, p = 0.03, BCa 95% CI = [0.109, more agreement with negative-scale items (e.g., the unethical use 1.236], with Hedges’ correction of 2.23. Their post-reveal stress of AI, the erroneous nature of AI, and its dangerous nature). A levels (M = 5.64, SD = 1.48) had no significant difference from more negative emotional valence of exposure is also related to less their pre-reveal stress levels, t (54) = 0.70, p = 0.49, BCa 95% agreement towards the positive-scale items (e.g., “AI systems can CI = [−0.164, 0.436], with Hedges’ correction of 1.17, reflecting perform better than humans,” “AI can provide new that even when the group realized that they were receiving support economic opportunities”). from an AI, the revelation of the true condition (or the activation Given that frequency, emotional valence, and immersion are of AI perception) did not affect their stress-level ratings as the all carried by emotions, the one-to-one relationships between other SQ ratings did. previous exposure and a particular subscale of attitudes reflect Participants regarded the helpfulness of the counseling chatbot the emotional attraction between individuals’ experiences and as neutral. The Told-Human group (M = 4.65, SD = 1.27) and the attitudes and also align with previous findings about negativity Told-AI group (M = 4.40, SD = 1.03) did not differ in perceiving salience (Robertson et al., 2023; Zollo et al., 2015). For example, the chatbot’s helpfulness, t (108) = 1.16, p = 0.25, BCa 95% the relationship between negative emotional valence and CI = [−0.161, 0.660], with Hedges’ correction of 1.16. disagreement towards positive-scale items (rather than agreement towards negative-scale items) could be explained by negativity salience and our inherent emotional tendency to oppose or strike 4 Discussion back at something we find “unright.” If the sample had some unpleasantly charged feelings in AI exposure, they could Although the knowledge and development of AI in Hong Kong be emotionally triggered by the positive statements about AI and are quite robust, the United States has a larger scale, more strike back by rating more negatively towards these pleasantly resources, and a faster pace of innovation for AI development charged statements, while also reassuring their pre-existing (Maslej et al., 2025). The difference in resource availability, beliefs about AI (Nickerson, 1998; Wason, 1960). It results in a investment, and industry presence may limit locals’ exposure to the more salient positive relationship between emotional valence and latest discussions and challenges surrounding AI. These explain the positive-scale attitudes. Since frequency and immersion also limited exposure to AI, neutral emotional valence, and moderate carry emotionally charged mechanisms (i.e., conditioned feelings immersion in our sample. Given the limited perceived personal about AI and the emotional bonding during immersive and social relevance, the sample had neutral attitudes towards AI experiences), the salience of negativity and emotional attraction and relatively low levels of AI anxieties. may explain their negative relationships with negative- Despite adequate validity, the minor deviations in the validity scale attitudes. results of AIAS from the original study suggested the need for While the overall sample showed that a stronger immersion further examination of item-level performance. Regarding the was related to significantly more unfavorable post-chat rather than validity deviations from the original studies of GAAIS, the 4-factor pre-chat positive-scale attitudes, exploratory analyses revealed this solution of the GAAIS found in this study suggests that the original effect was significant only in Gen Z. A person with low immersion structure may not fully replicate in our sample. The deviation in unpleasant experiences is less affected or restricted by might be attributable to the cultural differences in how attitudes pre-existing impressions about AI when receiving the chatbot’s towards AI manifest in Asians compared to Schepman and support since their attitude formations are less influenced by the Rodway’s (2020) UK sample. Future cross-cultural validation impact of immersion. In other words, they are more likely to studies may inform whether the GAAIS’s factor structure holds unfreeze their attitudes towards AI after receiving chatbot across cultures. counseling, resulting in more apparent attitude changes after the session. This is especially true for Gen Z. Given their fewer life experiences and exposure to the world than non-Gen Z, they 4.1 Emotionally charged statements and generally have more flexible opinions and attitudes. It aligns with negativity salience and is supported by previous studies on the effects of age on neuroplasticity (Hedden and Gabrieli, 2004), as well as the greater Consistent with the justifications using Pavlovian conditioning, adaptability and openness to change observed in Gen Z or younger fight-or-flight reactions to threats (Ekman, 2009), and the findings individuals (Fuchs et al., 2024; Visser and Krosnick, 1998). In about message internalization (Green and Brock, 2000; Valkenburg contrast, stronger immersion is associated with more rigid and Peter, 2013), the significant relationships between previous maintenance of pre-existing attitudes due to confirmation bias exposure and attitudes towards AI support the expectation that a (Nickerson, 1998; Wason, 1960), resulting in steady or unnoticeable higher frequency, more negative emotional valence and stronger changes in attitudes after receiving the chatbot’s support. Thus, Frontiers in Psychology 15 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 immersion has a stronger and more salient negative relationship and Peter, 2013), people who have more emotional investments in with post-chat rather than pre-chat positive-scale attitudes. exposures that illustrate AI as undesirable (e.g., movies that depict AI as a self-conscious destructive villain, or news about job replacement by AI) are more likely to fear humanoid AI configuration and concern 4.2 Urging for cautious AI developments about its increasingly sophisticated development in performing human abilities. Since configuration anxiety could signal the To the researchers’ knowledge, related studies investigating the formation of negative attitudes towards AI, we advocate for the relationship between exposure and AI anxiety have been conducted responsible development of AI through industry regulations and in organizational contexts (Elfar, 2025; Kong et al., 2021; Zhou et al., guidelines to alleviate public fears and concerns about AI’s 2024). This study addresses the gap by focusing on the general humanoid features. public and examining exposure dimensions to enrich broader discourse on the formation of AI anxieties. Consistent with related studies about AI awareness and AI anxiety, individuals’ frequency, 4.3 Necessitating the development of AI emotional valence, and immersion in exposure are each related to counseling attitude scale particular AI anxieties. Specifically, a higher frequency of AI exposure is associated with greater sociotechnical blindness, which As mentioned, SQ ratings towards the counseling chatbot only is particularly pronounced among non-Gen Z individuals. Aligning serve as indicators or partial reflections of attitudes towards AI with cognitive dissonance theory (Festinger, 1957) and related counseling. Given the absence of relationships between most studies on ambivalence and uncertainty (Buttlar et al., 2024; van general attitudes and AI counseling SQ ratings in our study, Harreveld et al., 2009), individuals with greater exposure to dual- developing an attitudinal scale for AI counseling is essential to perspective information about AI may experience uncertainty inform more sound investigations of the relationship and provide regarding the societal implications of AI-driven change. Indeed, the more complete reflection of attitudes towards AI counseling. insecurity of change has been a central concept in social psychology, Several subscales that measure support quality, counseling suggesting that humans are inherently conservative and prioritize accessibility, ethical considerations, user experience, and clients’ tradition over societal change (Jost, 2015). It is attributable to the autonomy when engaging with the chatbot would be examples to desire to sustain the comfortable state secured by collective interests, help predict public attitudes and the adoption of AI counseling in shared reality, and a sense of belonging, as well as the maintenance a more comprehensive manner. The development of such a scale of mental equilibrium (Cancino-Montecinos et al., 2020). Given could also enable practitioners in related fields to systematically that non-Gen Z individuals were not raised in a fully digitalized evaluate AI counseling and identify areas for improvement, thereby society like Gen Z, frequent exposure to dual-perspective better meeting clients’ needs. It could also inform the development information about AI could intensify this conservative human of guidelines to ensure the proper use of AI in counseling and nature and make them more insecure about AI’s advancement allow researchers to conduct comparative studies on attitudes in over time. various populations, perspectives (e.g., clients vs. counselors), and Nevertheless, reporting unpleasant incidents related to AI is cultures, to gain a better understanding of the societal implications necessary, as it prompts the need for regulations and remedial of AI counseling. strategies to address the issues presented. To minimize the negative Likewise, the nonsignificant relationships between AI anxieties impact of sociotechnical anxiety on people’s attitudes and adoption and SQ ratings suggest that the domains or conceptual frameworks behaviors, AI developers should take the lead in emphasizing that underpinning perceived support satisfaction towards counseling humans are always the masters of technology and social change. While chatbots may not be adequate to reflect complete attitudes towards AI developers are enthusiastic about the progressive advancements in AI, counseling. Another possible reason is that since the AIAS and GAAIS focusing on iterative improvements to existing systems, rather than were designed and considered AI in an all-in-one manner, they did rapid and radical deployment, may reduce failure rates and their not fully cover the features or conceptual frameworks of using AI in subsequent reporting. These considerations become especially vital the counseling context. Reviewing back to the relationship between when accounting for the negative relationship between emotional general attitudes and SQ ratings, a notable positive correlation is valence and learning anxiety found in this study, which aligns with observed only between post-chat positive-scale attitudes and SQ Elfar’s (2025) findings that high AI awareness could amplify ratings. The only significant result in the “post-chat” ratings is that the employees’ AI learning anxiety. The development of insecurity could support quality was assessed only after receiving the chatbot’s support. adversely affect people’s decisions or confidence in learning AI, The “positive-scale” items are more related to psychological subsequently hindering AI literacy. The relationship was particularly satisfaction, which directly shares a similar conceptual nature with noticeable to Gen Z since they are the generation more likely to learn well-being and support satisfaction in AI counseling (e.g., Q4 AI to keep pace with societal advancements, unlike non-Gen Z “Artificially intelligent systems can help people feel happier” and Q11 individuals who may already have established careers. It therefore “Artificial Intelligence can have positive impacts on people’s well-being”). underscores the need for gradual development and prioritizes Given the conceptual relevance, post-chat positive-scale attitudes are improvements on existing programs while allowing sufficient time for related to AI counseling SQ ratings. societal adaptation. In light of this, a more comprehensive examination of the Immersion in AI exposure is found to correlate positively with the relationship between AI anxieties and SQ ratings could be conducted development of AI configuration anxiety. Consistent with the findings by enriching the prospective AI counseling attitude scale with about message internalization (Green and Brock, 2000; Valkenburg conceptual frameworks related to AI anxiety (e.g., potential concerns Frontiers in Psychology 16 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 over role displacement in counseling, discomfort with communication The presentation difference of questions regarding stress levels and interaction with AI). The incorporation of these complementary and helpfulness projects the contradiction between implicit items could potentially yield more fruitful study results regarding effectiveness and explicit reservations about helpfulness. Perceived their relationship. stress levels were assessed before (i.e., initial stress level) and after the chat (i.e., post-chat stress levels), meaning that each stress level rating was separated by some time. Participants were unlikely to recall their 4.4 Emergence of perceptual fear in AI initial stress levels, nor did they have a “standard” in mind to project counseling their pre-existing beliefs about AI. In other words, they rated post- chat stress levels based on their true experience after the chat (and the Although support quality did not provide a complete picture of revelation). In contrast, the one-time item about perceived helpfulness people’s attitudes towards AI counseling, it provided information offered a more straightforward way to project pre-existing beliefs about individuals’ perceived support satisfaction regarding the because of its explicit presentation. After all, stress levels would working alliance between chatbots and humans. To the researchers’ be more reliable indicators of the counseling chatbot’s effectiveness in knowledge, there has been no prior attempt to investigate perceptual providing emotional outlets, given that its presentation is less fear or biased support quality appraisals towards counseling chatbots. contaminated by perceptual fear. Our study aligns with previous related studies about people’s mistrust of information from computers and algorithms (Dietvorst et al., 2015; Promberger and Baron, 2006). Our between-group results of SQ 4.6 Limitations and other future directions ratings demonstrated that people generally have biased appraisals towards AI in the counseling context. The perceived support Despite the adequate reliability and validity reflected by mean satisfaction of the Told-Human and Told-AI groups significantly inter-item correlations, the exposure subscales (i.e., frequency and differs even when both groups indeed received the same emotional valence) have not been properly validated. Findings about chatbot support. exposure should be interpreted as preliminary and replicated with In accordance with the confirmation bias theory, the validated measures in future work. As mentioned earlier, the nonsignificant within-group difference between the pre-and-post-chat demographics of the analyzed sample could not be compared with general attitudes and the less favorable support quality ratings towards those of the original sample, as the demographic data were not the “AI” label reflect the human propensity to maintain mental collected until the last set of questionnaires, and only valid responses equilibrium (Cancino-Montecinos et al., 2020), and people tend to were retained for data analyses. Future work should collect interpret or distort newly received information to reinforce demographic data at baseline and retain it to allow for the assessment pre-existing beliefs (Nickerson, 1998; Wason, 1960). In other words, of sample representativeness. the Told-AI group exhibited biased and more negative appraisals Since this study only used perceived support qualities as towards AI performance than its actual capability. Perceptual fear was indicators of attitudes toward AI counseling, it may not also observed when the Told-Human group rated support quality sufficiently provide a complete picture of public attitudes toward more negatively after being revealed the true condition. It implies that AI counseling. Furthermore, given that the data were collected even when they knew that the support was the same regardless of the from local Chinese in Hong Kong, the results may only reflect the revelation, the effect of perceptual fear on perceived experiences local context. Future studies could consider conducting cross- was salient. cultural examinations of perceptual fear in AI counseling and While the emergence of perceptual fear introduces challenges in developing a comprehensive attitudinal scale to measure public obtaining unbiased support quality ratings within the AI counseling attitudes toward AI counseling. context, concealing AI support is unacceptable as clients have the right Considering the typical working hours (i.e., 8–10 h) for full- to know the kind of support they would receive and “who” they would time workers in Hong Kong, the time limit for participants to share the issues with. Future studies may investigate the impact of complete each survey was set at 12 h after the survey link was perceptual fear on perceived support satisfaction in countries that issued to provide sufficient time for completion while minimizing deploy AI counseling support tools (e.g., the United States), so as to the dropout rate. The 12-h latency, however, may reduce accuracy inform potential strategies for mitigating perceptual fear in the context stemming from short-term memory effects. Meanwhile, since this of AI counseling. study focuses on chatbot counseling, where prospective clients typically interact with it in real-life settings, experimenting in a laboratory may not yield more representative results. To facilitate 4.5 Implicit effectiveness and explicit memory retention, future studies could impose a stricter time limit reservations about helpfulness for ratings; however, caution should be exercised to minimize drop-out rates. They could also recruit a larger sample size to While people generally showed more biased and negative strengthen the generalizability of results. evaluations of AI, the significant reduction in stress levels is consistent The system message given to the counseling chatbot in this with previous findings on the effectiveness of counseling chatbots in study generally employed the CBT approach due to its well- improving users’ mood (Fitzpatrick et al., 2017; Inkster et al., 2018), researched effectiveness and its structured nature that is which reflects AI’s capability to provide emotional outlets. However, compatible with AI. Future explorations could utilize AI to people had reservations about its helpfulness, which may be due to the deliver counseling techniques that involve more dynamics and blockage by their pre-existing beliefs. variations (e.g., psychodynamic and humanistic approaches) to Frontiers in Psychology 17 frontiersin.org Keung and So 10.3389/fpsyg.2025.1538387 examine whether similar findings can be obtained. Further Author contributions explorations are encouraged to utilize different support quality measures to examine the existence of perceptual fear in AI WK: Conceptualization, Data curation, Formal analysis, counseling, as well as to conduct sentiment and thematic analysis Investigation, Methodology, Project administration, Visualization, of the conversations to compare the emotional tone and choice of Writing – original draft, Writing – review & editing. TS: Software, words used in the Told-Human and Told-AI groups, to inform Supervision, Writing – review & editing. about their relative willingness or openness to disclose with a “human” or “AI.” Fu