Cigarette smoking is one of the most preventable causes of premature death and illness [1], and supporting smokers in their attempts to quit is a public health priority. While the use of cessation support (eg, pharmacological treatment and behavioral support) can effectively increase the successful quitting rate [2], such services are facing substantial staff and resource shortages [3,4], which have been aggravated by the COVID-19 pandemic. Moreover, people tend to express resistance to using professional support when it comes to quitting smoking [5]. To combat these challenges, digital tools that can complement traditional cessation services have become increasingly popular. Conversational agents, or chatbots (ie, computer programs that use artificial intelligence to simulate a conversation with users through natural language [6]), are one of the recent innovations that bring several benefits: accessibility, scalability, personalization, and synchronicity, to name a few [7]. These opportunities have sparked interest to test the use of chatbots in health care settings, and initial effectiveness has been found in various domains, such as healthy lifestyle promotion [8], sexual health [9], mental health [10], and smoking cessation [11]. However, it is still unclear when and how chatbots are acceptable and effective. Some studies in human-computer interaction have shown that the positive effect of chatbots tends to decrease over time [12], and a few barriers that hinder human-chatbot interaction have been identified, including a lack of engagement caused by counseling styles that users do not prefer [13]. As long-term support is essential for smoking cessation, which usually requires sustained effort, and in view of the one-to-one nature of chatbots that allows for individualized conversation, it is important to find the optimal counseling style to overcome potential resistance to effectively motivate smoking cessation.

Within the field of individualized smoking cessation counseling, a chatbot can use various counseling styles, such as confrontational counseling (CC) and motivational interviewing (MI). CC originated in alcoholism treatment and emphasizes a hard-hitting and directive style intended to break through patients’ defense mechanisms [14]. CC counselors are trained to confront patients with the consequences of their unhealthy behaviors to counter self-exempting beliefs and use direct, unsolicited advice to increase patients’ risk perception [15]. In the field of smoking cessation, CC techniques include direct education about the risks associated with smoking, challenging smokers’ minimization or denial of the problem, and urging for abstinence by providing a quitting plan [16-18]. A typical example situation is the counselor coercing the client to face up to reality by emphasizing their problematic smoking behavior and the associated risks [14]. However, this does not necessarily entail an aggressive approach but rather attempts to raise clients’ awareness of the likely severe consequences of their smoking behavior [19]. Indeed, contrary to the common assumption that confrontation leads to denial [15], CC has been found to be predictive of higher patient involvement and higher long-term abstinence [20,21]. For several decades, CC has been regarded as the only way to get patients to listen [14], and there is positive evidence of health care providers using directive confrontational tactics to advise smokers to quit [22]. However, recent research suggests that the effectiveness and acceptability of CC depends on the counselor’s legitimacy and a rather empathetic manner [23], whereas resistance may escalate if the client feels threatened by the confrontational tone [24].

In response to the potential resistance, an alternative approach has been developed where patients’ autonomy and intrinsic motivation are highlighted [14,25]. Miller and Rollnick [25] introduced MI, a nonconfrontational counseling style for eliciting behavior change by helping clients explore and resolve ambivalence. The strategies of MI are more subtle than coercive, prioritizing encouragement over contention, with the ultimate goal of eliciting the client’s intrinsic motivation rather than imposing or desiring behavior change [25,26]. MI counselors use the principles of expressing empathy, avoiding arguing and confrontation, and supporting the individual’s self-efficacy. A typical example is “What you choose to do is up to you. My role is to help you figure out what you want to do and support you in accomplishing your goals” [27]. In the context of smoking cessation, MI counselors use techniques such as asking open-ended questions to elicit personally relevant reasons to quit, reflecting on the client’s words to encourage disclosure, affirming the client’s expressions toward quitting, and summarizing to help the client have a better overview of their thoughts [25,28]. The MI counselor assists the client in realizing the discrepancy between their values and their behaviors, thereby eliciting intrinsic motivation to resolve this discrepancy by changing their behavior [29,30].

Over the past decades, a rich amount of research on the effects of MI and CC on smoking cessation and other health behaviors has spawned, yielding mixed and inconclusive results. While CC has been found effective in increasing smokers’ risk perceptions and, in some cases, led to smoking cessation in the long run [20,31], there is also evidence suggesting that CC may lead to defense mechanisms among the clients (eg, denial and dismissal of the information) [32]. Similar inconclusiveness has been observed regarding the effectiveness of MI; while successful with other behavior domains (eg, alcohol use and physical exercise), the treatment tends to have weaker effects on smoking cessation [33,34]. Despite the mixed evidence regarding the 2 styles, modern addiction treatment research believes that MI is more likely to be successful than CC [14]. In general, MI-style interventions have been found to motivate a larger proportion of individuals than CC-style interventions, making them a more applicable and efficacious approach [14,35]. Moreover, the effectiveness of CC depends on the legitimacy and authority of the counselor [36]. In the context of chatbot-delivered interventions, CC might not be as effective as chatbots are usually not regarded as a medical authority [37]. On a broader note, it has been suggested that people prefer chatbots that offer emotional support (ie, MI) more than chatbots that predominantly offer informational support (ie, CC) [38]. In summary, MI seems to be a better approach for a smoking cessation chatbot.

However, it remains unclear whether the hypothesized advantage of MI over CC can be experimentally sustained in a chatbot setting. To date, only a few studies in human-delivered counseling have directly compared the 2 styles and found mixed results such that MI (vs CC) did not significantly reduce drinking and drug use but yielded less resistance and higher client satisfaction [14,18]. Moreover, no study has explicitly compared the 2 styles in a chatbot setting, and it is yet unknown whether the effect of MI is also applicable in automated settings. Human-computer interaction research in health care has emphasized the chatbots’ relational strategies to improve their effectiveness and acceptability [39,40], which are keenly matched to the principles of MI [25]. However, despite the theoretical advantages, the empirical evidence of automated MI is inconclusive [41]. Such insignificance of MI could be explained by the methodological disparities in the line of research: the comparison groups varied considerably, and the robustness of MI cannot be concluded. In a recent review of technology-delivered MI interventions [41], various studies compared MI with other types of interventions (eg, self-help booklet and assessment only) but not with an alternative counseling style (eg, CC). Therefore, the effectiveness of MI as a chatbot counseling style cannot be ascertained. In fact, regardless of the therapeutic counseling style, people might find a chatbot acceptable and even persuasive when presented nonintrusively [42], which, therefore, shadows the effect of chatbot counseling styles. Indeed, in an earlier study [43], an MI-style chatbot was compared with a neutral-style chatbot in motivating smokers to quit, expecting MI to generate a higher intention to quit and better therapeutic experience among smokers. However, no significant difference between the MI style and the neutral style was found. To follow up on previous research and further understand the role of the chatbot and better ascertain the usefulness of MI as a chatbot counseling style, this study compared the MI and CC therapeutic counseling styles in a chatbot setting. To sum up, this study hypothesized that an MI-style chatbot, compared to a CC-style chatbot, results in higher self-efficacy, stronger motivation to quit, more engagement, a stronger therapeutic alliance, more perceived empathy, and higher interaction satisfaction.

Moreover, to have a more comprehensive picture of whether and how an MI (vs CC) chatbot works and understand how users experience the chatbot interaction, we explored users’ perspectives on the chatbots through semistructured interviews. Specifically, we examined the following aspects: (1) What is the overall user experience with the smoking cessation chatbots? (2) How does the chatbot influence smokers’ perceptions of smoking and intention to quit? (3) How can we improve the chatbots for future use?

This mixed methods study combined a quantitative part using a web-based experiment and a qualitative part of semistructured interviews. For the experiment, participants were randomly assigned to interact with a chatbot using an MI style or a chatbot in a CC style. Among a smaller set of participants, semistructured interviews were conducted after the chatbot interaction to gain further insights into user experience.

Ethics approval was obtained from the Research Ethics and Data Management Committee of the Tilburg School of Humanities and Digital Sciences (identification code REDC 2021.18ab), and the study was conducted in compliance with the ethical and data management regulations of the school. Informed consent was obtained from participants via the Qualtrics (Qualtrics International Inc) web form. Participants were recruited via the participant pools from the Tilburg School of Humanities and Digital Sciences and Tilburg School of Social and Behavioral Sciences and were compensated with credits. The study design, raw materials, and analysis plan were preregistered at the Open Science Framework.

Power calculations were conducted for multivariate ANOVA (MANOVA) using G*Power (version 3.1) [44], indicating that a sample size of 226 participants is adequate to uncover medium effects (effect size: ƒ²=0.0625; power=0.8) in accordance with previous meta-analyses on the effects of MI on smoking cessation [45]. Eligible participants were smokers (ie, had smoked at least one cigarette in the week before participation) aged ≥18 years with a competent proficiency in English reading and writing.

In the CONSORT (Consolidated Standards of Reporting Trials) flow diagram in Figure 1, the study procedure is visualized. Upon starting the study, participants completed the pretest questionnaire assessing demographics and baseline motivation to quit smoking, after which they were randomly assigned to interact with either the MI or the CC chatbot for 2 consecutive sessions, simulating a typical intake session and a first consultation session in smoking cessation interventions [46,47]. The randomization was double blinded and was carried out automatically by the Qualtrics software. In the first session, the chatbots carried out a conversational assessment of the participants’ smoking behavior and initial motivation and barriers to quitting smoking. After approximately 5 to 10 minutes, which resembles the waiting room setting, the second session started, in which the chatbot discussed with participants their previous quit attempts, aimed to strengthen their self-efficacy, and encouraged the participants to form future quit plans. Relevant outcome measures were assessed after each session. Upon completion, all participants were debriefed. In addition, interview invitations were sent out at random to participants in both conditions, and the final interview sample consisted of 16 participants, 8 (50%) from each condition.

Demographics and smoking-related baseline information were collected at the pretest time point. Age and gender were assessed using single items. Baseline motivation to quit was measured using the Contemplation Ladder [53], where participants indicated their motivational status ranging from 0 (no thoughts on quitting) to 10 (taking action to quit). Level of nicotine dependence was measured using the Fagerström Test for Nicotine Dependence [54], and average daily cigarette consumption was measured using a single item.

For outcome measures, we assessed both cessation-related variables and user experience–related variables, and all variables were measured both after the first session (T1) and after the second session (T2). All scale variables were measured using 5-point scales. The measurement scheme followed that of the earlier study [43] to facilitate comparison.

Intention to quit was measured using the Contemplation Ladder [53]. Participants indicated where they identified themselves on a single item, with responses ranging from 0 (no thoughts on quitting) to 10 (taking action to quit).

Self-efficacy was assessed using the Smoking Self-Efficacy Questionnaire [55]. On a 5-point scale, participants indicated whether they thought they would be able to refrain from smoking in various difficult situations. An example question is as follows: “how sure are you that you could refrain from smoking when you feel nervous.”

Engagement with the chatbot interaction was assessed using 9 items from subscales of the short form of the User Engagement Scale [56]. The esthetic appeal subscale was removed as this study focused on the communication process instead of the interface design. An example item is as follows: “I was absorbed in this experience.” An additional question was included asking about participants’ endorsement for future use.

Therapeutic alliance was measured using the Working Alliance Inventory–Short Revised [57], a 12-item self-report measure used to assess the relationship between participants and the chatbot. An example item is as follows: “We agree on what is important for me to work on.”

Perceived empathy was measured using a 3-item (eg, Roby seemed to understand me) scale based on research on interpersonal communication by Rubin and Martin [58].

Interaction satisfaction was measured using 5 items on a 5-point scale ranging from “Not at all” to “Very much” following studies on telephone smoking cessation counseling [59]. An example item is as follows: “how was your counsellor in terms of being a good listener?”

As a manipulation check question, perception of MI was measured using an 8-item scale based on the Client Evaluation of Motivational Interviewing scale [60]. Example items are “Roby argued with you to change your behavior” and “Roby helped you feel confident in your ability to change your behavior.”

Experiment data were analyzed using R (version 4.2.1; R Foundation for Statistical Computing). Independent-sample 2-tailed t tests and a chi-square test were conducted to check for equal distribution of background variables (ie, gender, age, nicotine dependence level, and baseline motivation to quit) across conditions. Variables that were not equally distributed across conditions were included as covariates in hypothesis testing.

To test the preregistered hypothesis regarding the effect of MI chatbot conversations on the outcomes, a 1-way MANOVA was performed using the 6 outcome variables (ie, self-efficacy, intention to quit, engagement, therapeutic alliance, perceived empathy, and interaction satisfaction). Data measured at T2 were used for this hypothesis.

Qualitative data were gathered at the postmeasurement time point through semistructured video interviews with 16 participants, 8 (50%) each randomly selected from the 2 conditions. The interviews lasted 20 to 30 minutes each and were conducted by the first author, who was familiar with the chatbot content. Before study commencement, an interview guide was developed by the author team and pilot-tested. The interview topics were guided by the theoretical framework of acceptability of health care interventions [61], which consists of several constructs (eg, affective attitude, burden, intervention coherence, and perceived effectiveness) that closely relate to user experience with health care interventions.

The interviews were audio recorded and transcribed. The transcripts were analyzed using thematic analysis [62]. We adopted a hybrid approach combining deductive and inductive coding [63] in which coding was mainly informed by themes that could potentially answer the research questions and inform chatbot improvement but also allowed for novel codes to emerge for additional insights. A pilot coding round was carried out by 2 independent reviewers, and consensus regarding the codes was reached through discussion. Coding was performed in an iterative process over multiple reads of the transcripts. Multiple group meetings were organized aiming to identify potential biases in the coding and interpretation process. One author (LH) carried out the coding of the remaining interviews. The analysis procedure was performed using the ATLAS.ti (ATLAS.ti Scientific Software Development GmbH) program.

In this mixed methods study, we aimed to investigate the effectiveness and user experience of MI strategies delivered by a chatbot. Results showed support for our hypothesis regarding the advantage of the MI chatbot in terms of user experience (ie, engagement, therapeutic alliance, perceived empathy, and interaction satisfaction), but no significant effect was found for the cessation-related outcomes (ie, intention to quit and self-efficacy). Importantly, exploratory analyses showed increases in participants’ intention to quit, self-efficacy, and therapeutic alliance over the sessions, suggesting the usefulness of chatbots for this purpose. These increases were more profound in the MI condition (except for therapeutic alliance, for which the increase over time did not differ between conditions), suggesting the potential of MI in the long run. Finally, participants reported an overall decrease in their engagement with the chatbot. Qualitative data supported and expanded the experiment findings. We identified constructs that relate to users’ prospective acceptability, timely experience, and retrospective acceptability of our smoking cessation chatbot, with several insights applicable to health chatbots in general.

Our findings regarding user experience are in line with those of previous research not involving chatbots demonstrating clients’ preferences for an empathetic and collaborative approach. Previous research in health counseling has highlighted patients’ preferences regarding an empathic and collaborative approach [14,18], and our findings echo such preferences and extend their value to the chatbot setting. Our qualitative results provided further possible explanations for this. Participants in the MI condition mentioned more frequently than participants in the CC condition that they felt understood by the chatbot and that they were more engaged in the conversation. In the CC condition, participants more frequently felt coerced and confronted by the chatbot, and they indicated that they would like more personal conversation and that the provided information was not sufficient. These findings show that participants value the relational aspect of the chatbot, suggesting that it is essential for the chatbot to have a person-centered approach and have a supportive attitude. Moreover, several participants mentioned having previous expectations of an authoritative and confrontational figure sending generic information, whereas the MI chatbot exceeded their expectations with its personalized, understanding, and empathetic approach. Client expectations regarding treatment are known as a key factor in counseling outcomes and satisfaction [65]. Expectancy violation theory suggests that positive expectation violation (eg, the chatbot was more personalized than expected) predicts more favorable outcomes than negative expectation confirmation (eg, the chatbot was confrontational as expected) [66]. Indeed, despite recent research favoring MI, many addiction counselors may still use confrontational techniques. Additionally, smokers often face criticism from their social environment, which may influence their expectations of how a chatbot will interact with them [16,51]. In other words, the previous expectation of a CC-like chatbot interaction may account for the positive user experience with the MI chatbot.

In terms of the motivating effect on smokers’ intention to quit and self-efficacy, our results did not demonstrate significant differences between the 2 chatbot styles. The MI chatbot did not evoke more motivation or self-efficacy than the CC chatbot. One plausible explanation for why we did not find differences regarding the different styles could be that our participants were recruited with established motivation at baseline. On average, participants in both conditions placed themselves on the baseline contemplation ladder between “I think I should quit but not quite ready” and “I am starting to think about how to change my smoking patterns.” Smokers at this stage of change might appreciate both encouragement and concrete guidance on quitting [67]. In other words, participants may value the intrinsic motivation-evoking aspect of MI, whereas the direct advice in CC is perceived as useful at the same time. Together, these effects may have resulted in nonsignificant differences in motivating intention to quit between the 2 styles. Another interpretation stems from the particular chatbot setting. While we hypothesized the MI chatbot to be helpful in evoking motivation and encouraging personally achievable quit plans, its effectiveness might depend on clients’ actual perception of it. The fidelity of MI requires true understanding and empathy, and it remains unclear whether the chatbot encompasses these qualities. Research in human-computer interaction has suggested that inaccurate artificial empathy might be less favorable than no empathy [68,69]. This is supported by our interview findings such that, despite the clearly appreciated nonjudgmental and supportive tone of the MI chatbot, people preferred discussing emotional experiences with real humans because chatbots do not have emotional capacities. On the other hand, while CC was expected to have less positive effects, using a chatbot as the portrayed counselor might have mitigated the anticipated resistance. As noted by both our interview participants and previous research, health chatbots are predominantly perceived as an instrumental tool that answers questions and provides information [70], and therefore, people might value the informational aspect more than the therapeutic aspect. However, it is important to highlight that, regardless of the condition, participants overall reported a small yet significant increase in their motivation to quit and self-efficacy after the chatbot interaction. This is consistent with previous research suggesting that minimal conversation about smoking cessation can affect quitting intentions and behaviors [71], and chatbots might be particularly useful due to their high accessibility. It is conceivable that the use of a chatbot alone is effective, and the specific communication strategies used may not have a major impact on the outcomes [43]. To better realize their potential, future research is needed to understand the impact of different automatic conversation strategies and the role of users’ communication preferences regarding chatbots and design chatbots that combine effective strategies from different counseling approaches.

Another notable finding is that people’s engagement with the chatbot decreased over time. Despite the fact that our experimental setting consisted of only 2 sessions, this provides a stepping stone for the long-term use of chatbots, which is particularly important for smoking cessation, which requires sustained effort. Research in human-computer interaction shows that user engagement tends to decline over time [72-74]. In addition to the novelty effect (ie, people engage with new technology due to curiosity, which declines gradually [75]) that has been proposed to explain declining engagement [12], our study found that users’ technical experience was an active determinant in the engagement process. The most frequently mentioned reason for participants to disengage from the chatbot interaction was the encountered technical errors (eg, chatbot misunderstanding and repairing). Being able to input free text was one of the aspects that participants found engaging. This provides a plausible explanation for the finding that the MI chatbot was perceived as more engaging than the CC chatbot such that the open-ended questions asked by the MI chatbot offered users active involvement in the conversation. To increase people’s engagement with the chatbot and the intervention, future research is needed to identify more factors that influence the engagement process. One area worthy of investigation relates to the automatic detection of engagement or disengagement from user utterances so that the chatbot can respond in time to keep the users engaged. For example, He et al [76] explored textual cues indicative of user engagement and suggested using cognitive strategies (eg, quizzes and reflective questions) to keep the user engaged.

A number of methodological limitations should be borne in mind when interpreting the results. First, the length of the interaction was fairly short. We aimed to simulate a brief intake session and a first counseling session, and our results were in line with research on the positive effects of brief interventions for smoking cessation [77]. Even though we found a significant increase in smokers’ motivation and self-efficacy to quit, this increase was relatively small, which is not unexpected given the briefness of the interaction. Our exploratory results suggest that the small but positive effects increased over time. The promising outlook calls for studies that involve prolonged use to seek more meaningful changes and investigate whether the effects can be sustained in the longer term. We believe that chatbots are well suited for addressing this long-term objective as they facilitate extended and personalized interactions, and we plan to delve deeper into this potential in a follow-up longitudinal study. In addition, we measured intention rather than actual cessation as a primary outcome, which provided initial support for the usefulness of the chatbot interaction. To further this line of research, behavioral outcomes are needed to ascertain the effectiveness of chatbot interventions. Moreover, objective use data (eg, interaction length and sentence length) can be used to complement self-reported engagement. It is recommended that future research include more objective instruments such as biochemical and behavioral measures to validate the effects of chatbot interaction. In addition, although our sample represents the main user group of new technologies such as chatbots, their smoking behaviors might not be representative of the general smoking population. Their familiarity with technology might have also led to results that are less applicable to other user groups. Considering the potential impact of demographics and smoking behaviors, it is important to design chatbots that can cater to a broader range of populations.

Our findings, coupled with previous research, suggest that chatbots have the potential to motivate smokers to quit, and equipping them with MI skills could improve user experience. However, the insignificant results on cessation-related outcomes call for more research to further identify strategies that can effectively motivate cessation. For example, previous research has attempted to understand the working mechanism of MI in an automated setting by comparing different types of open-ended questions and reflections and suggested the advantage of incorporating multiple-choice questions particularly in an automated setting, expanding on the general guideline of using open-ended questions in traditional MI [28,78]. Therefore, it is important to disentangle the MI components and translate active ingredients to a chatbot setting. Moreover, our findings suggest that the directive informational aspect of the CC chatbot was appreciated by the participants, and future research is encouraged to combine such aspects with the nonjudgmental spirit of MI to capture the best of both worlds. In addition, in view of the person-centered approach and the potential of personalization, it is important to explore individual factors that can influence people’s preferences and experience with the chatbots. For instance, previous research has found that users’ need for autonomy and their self-efficacy was influential in their responses to an MI or a CC chatbot [79], highlighting the importance of addressing users’ unique needs in designing future chatbot interventions. As noted in the qualitative findings, technology-delivered counseling has both positive (eg, easy access and facilitated disclosure) and negative (eg, lack of true understanding) effects. Considering the interpersonal nature of health counseling, it is essential to understand the role of the human touch in chatbot-delivered interventions. Our study provided some initial insights into this, such as that the chatbots facilitated health-related disclosure but people are skeptical of their ability to understand human feelings. Future work is needed to further understand how people perceive chatbots in an interpersonal setting; for example, how do people view the relational communication delivered by a chatbot? How do people disclose their emotions and understand chatbots’ responses? This line of research will help improve the design of chatbot-delivered MI and health counseling in general.

On a practical note, researchers and practitioners should be cautious in humanizing chatbots. Even though research has pointed to the potential of empathetic agents [40], overly imitating human emotions could result in an uncanny feeling [80]. Our findings suggest that, while humanlikeness was generally favored, participants felt hesitant discussing emotions with the chatbot when they were aware of its mindless nature. To best equip chatbots with social cues, future work needs to understand users’ perceptions of different facets of emotions. For example, cognitive empathy (ie, understanding others’ perspectives and emotions) from a chatbot is appreciated, whereas affective empathy (ie, emotional response to others’ emotions) induces perceived eeriness [38]. Moreover, we argue that chatbots should be considered a supplementary tool rather than a replacement for human caregivers, especially in health-related contexts where inaccurate and invalid information can have a harmful impact on people’s well-being. Several of our interview participants suggested that the chatbot should be a standby conversational tool that provides timely support. Functions supporting long-term use (eg, quitting progress report) should be considered as well.

An important implication of this study is that previous expectations play a crucial role in people’s experiences with and responses to chatbot interventions. While people used to have expectations of a customer service–like generic chatbot, such expectations might be drastically changed by recent rapid technological advancements. In particular, ChatGPT, one of the latest large language models, has captured the attention of millions since its release. It might reshape people’s expectations of chatbots with its large amount of knowledge and natural text generation and its capability to respond to any conceivable input. However, it should be made clear to the public that such technologies should not be regarded as a replacement for medical authorities and are not ideal for long-term support as it requires personalized care and tailored interventions, whereas ChatGPT’s capabilities are limited to the general information that it was trained on. In addition, it is well established that chatbots relying on large language models may give socially desirable answers, may suffer from undesired biases, and are not necessarily truthful [81]. Despite the extensive amount of information available, such technologies simply answer to requests and often overlook user background and conversational contexts. We are aiming to develop chatbots that can understand and engage with users in the long term while ensuring accuracy and safety. Health care chatbots need to balance safety measures (eg, using prescripted utterances to avoid harmful content) with engagement features (eg, flexibility and variety in texts). Technologies such as ChatGPT could bring additional support by providing coherently written information, but human control should be involved to ensure safe, responsible, and accountable use.

This study adds to the body of research regarding the effectiveness and user experience of chatbots using 2 different counseling styles. The results point in the direction that chatbots are a promising tool in motivating cessation and the use of MI can improve user experience. Several constructs such as affective attitude and engagement were identified to understand users’ previous expectations, timely experience, and retrospective acceptability of the chatbot interaction. These findings highlight the potential of chatbots for smoking cessation and suggest a few avenues for future research.