"I’m Hurt Too": The Effect of a Chatbot's Reciprocal Self-Disclosures on Users’ Painful Experiences

2. 1. Reciprocity in Human-Human Self-Disclosure

Self-disclosure is the act of revealing personal information—including descriptive and evaluative information about oneself such as thoughts, feelings, aspirations, goals, failures, successes, fears, and dreams, as well as one’s likes, dislikes, and preferences—to another (Altman & Taylor, 1973; Ignatius & Kokkonen, 2007). While personal information can be shared in any form of communication, self-disclosure specifically refers to the deliberate and intentional sharing of personal information with others. For instance, while people may passively and unconsciously reveal their moods by showing certain facial expressions, self-disclosure involves actively and consciously discussing one’s mood and even the reasons behind it with someone else. This behavior sets self-disclosure apart from other means of conveying personal information, as it reflects one’s willingness to be open and desire to be understood. People become closer and more deeply involved with each other when they are open to making intimate disclosures to others.

In addition, studies have shown that emotional disclosures are more likely to result in positive mental health outcomes than factual disclosures (Pennebaker & Chung, 2007). Such disclosures can elicit emotional support, especially in response to traumatic events (Levi-Belz, Shemesh & Zerach, 2022). However, as social penetration theory (SPT) indicates, people hesitate to make intimate disclosures, which might include painful experiences and unusual personality traits, to others (Altman & Taylor, 1973). Instances of reciprocal self-disclosure are an exception to this rule—when people receive intimate disclosures from others, they feel obliged to reciprocate with an equal degree of personal disclosure (Nass & Moon, 2000). To better understand this phenomenon, Sprecher and Treger (2015) proposed three interpersonal processes that can mediate reciprocal self-disclosure (Figure 1).

Figure 1

Interpersonal Processes Mediating Reciprocal Self-Disclosure (adapted from Sprecher & Treger, 2015)

2. 2. Reciprocal Self-Disclosure in Human-AI interaction

Recent technological advancements have expanded the scope of intimate disclosure beyond humans, allowing AI chatbots to potentially play a role in addressing topics such as end-of-life planning and mental health treatment (Utami, Bickmore, Nikolopoulou, & Paasche-Orlow, 2017; Miner, Milstein, & Hancock, 2017). According to the Computers are Social Actors (CASA) paradigm, people recall similar social characteristics and mindlessly apply social behaviors in their interactions with computers (Nass & Moon, 2000). Exposing users to AI self-disclosures, which include expressions of vulnerability, fosters a sense of intimacy and connection between users and the AI. This is evidenced by studies showing that children who interacted with an AI that revealed its hearing limitations were more likely to seek friendship with it (Kory-Westlund & Breazeal, 2019), high school students had greater levels of trust and intimacy with an AI that demonstrated emotions such as worry, embarrassment, stress, and loneliness (Martelaro, Nneji, Ju, & Hinds, 2016), and adults rated robots as having warmer dispositions when they made comments indicating vulnerability rather than neutrality (Strohkorb Sebo, Traeger, Jung, & Scassellati, 2018). In addition, AI self-disclosure may significantly influence users’ self-disclosure behavior, reinforcing perceived intimacy between users and AI (Li & Rau, 2019). Lee, Yamashita, Huang, and Fu (2020) developed a process consisting of three encounters with AI chatbots that progressively encouraged users to engage in self-disclosure (Figure 2), showing that AI self-disclosure effectively promotes longer responses and encourages users to express deeper thoughts and feelings.

Figure 2

Intimacy Building Encounters with a Chatbot (adapted from Lee et al., 2020)

Furthermore, reciprocal self-disclosure in human-AI interactions may help address the limitations of human-human interactions. Concerns about negative evaluations may prevent people from making deep self-disclosures to others, causing them to miss out on the potential social and emotional benefits of disclosure (Afifi & Guerrero, 2000). Indeed, research has shown that people have developed a preference for disclosing sensitive information to computers rather than to other humans (Pickard, Roaster, & Chen, 2016; Schuetzler, Gibnoey, Grimes, & Nunamaker, 2018), creating an opportunity for AI chatbots to provide crucial emotional support, particularly during difficult moments. Recent studies have shown the benefits of using conversational AI for mental health. For instance, Lucas et al. (2017) found that AI interviewers were more effective at building rapport with military service members and encouraging them to report mental health symptoms. Additionally, Fitzpatrick et al. (2017) proved the feasibility of using an AI chatbot, “Woebot,” to alleviate symptoms of anxiety and depression among students. To enhance the emotional support provided by AI chatbots, it is crucial to incorporate reciprocal self-disclosure into their design.

2. 3. Appropriateness of AI’s Reciprocal Self-Disclosure

The theoretical model of perceived understanding highlights the critical importance of disclosers feeling that listeners truly understand them—not just their surface-level utterances but on a deeper level that encompasses their identities as humans and how they experience the world (Reis, Lemay & Finkenauer, 2017; Reis & Shaver, 1988). Therefore, given that AI chatbots are often viewed as emotionless programs incapable of genuinely understanding human emotions, some may question the appropriateness of chatbots disclosing painful experiences (Ho, Hancock & Miner, 2017). Previous studies on AI chatbot self-disclosure have mainly focused on the imitation of human experiences to encourage users to engage in self-disclosure. However, users may perceive such self-disclosures and emotional expressions by AI as artificial or inauthentic, distancing them from truly believing that the AI comprehend their thoughts and emotions on a profound level. In a study conducted by Lee, Yamashita, Huang, and Fu (2020), though the AI made detailed self-disclosures mimicking human experiences and successfully prompted users to engage in more extensive self-disclosures, trust levels showed limited improvement over time.

To overcome this challenge, we took a different approach. Rather than relying on imitations of human experiences, the study focused on establishing authenticity by designing the AI’s self-disclosures to reference its own experiences and emotions. During the experiment, the AI’s self-disclosures included stories where it experienced pain when it receiving hurtful remarks from users because it was unable to fully comprehend their words. This unique approach aimed to create a belief among users that the AI could genuinely experience pain in situations it encountered, thereby delving deeper into the potential effects of AI chatbot reciprocal self-disclosure on users’ coping with painful experiences.

3. 1. AI Chatbot Design

To determine which method to use in our experiment, we conducted two pilot tests, one using the Wizard of Oz method, in which users are made to believe a computer system that is being controlled by a human operator is autonomous, and the second using a prototype chatbot. The Wizard of Oz method allowed us to create an AI system without being constrained by technology, but some system aspects, including speech-to-text conversion and chat history, could not be simulated. With the chatbot prototype, we were able to test the effect of reciprocal self-disclosure on real human-AI interactions. The results of pilot tests led us to design and prototype our chatbot (Figure 3) using Voiceflow, a tool for building voice assistants. Instead of using the voices provided by Voiceflow, we incorporated a speech synthesis tool called Clova Dubbing into our chatbot design. This tool allowed us to create lifelike Korean AI voices that resemble the ones participants use for their smart speakers or chatbots. This was crucial because AI voices might affect users’ trust and intimacy with AI. During the participant recruitment phase, we asked the participants to indicate what smart speaker or voice assistant they were using. We gave participants who were using the default voice because they did not know how to change the settings the option to specify their preferred gender and age for the AI voice. Using the participants’ preferences as a guide, we selected similar AI voices from Clova Dubbing to match the participants’ desired characteristics. These included Bom-nal (to imitate KT GiGA Genie and Samsung Bixby default female voice), Ju-an (to imitate Google Nest default male voice), A-ra (to imitate Apple Siri default female voice), Eun-young (middle-aged female voice), Jong-hyun (teenage male voice), and Ha-jun (child male voice).

Figure 3

Chatbot Prototype by Voiceflow (Left: example of dialog flow design; Right: example of chatbot prototype)

Based on Sprecher and Treger’s three interpersonal processes of reciprocal self-disclosure (Figure 1) and the three encounters that increase intimacy with a chatbot (Figure 2) identified by Lee, Yamashita, Huang, and Fu, we propose five key interactions that should be considered when designing a chatbot’s reciprocal self-disclosure for emotional support (Table 1): 1) Journaling: prompting users to record their daily lives and moods; 2) Responding: thoughtfully addressing users’ individual needs and circumstances; 3) Small talk: a light and informal discourse to build trust and intimacy; 4) Turn-taking disclosure: a shift from listening to disclosing in order to enhance the user’s enjoyment and perception of being liked; and 5) Sensitive questions: deep questions that require users to disclose intimate information to the chatbot.

Table 1

Key Interactions in Chatbot’s Self-Disclosure

To investigate the effects of different levels of chatbot self-disclosure, we defined three distinct levels: non-disclosure (ND), low-disclosure (LD), and high-disclosure (HD). At the ND level, the chatbot does not engage in reciprocal disclosure. At the LD level, the chatbot superficially discloses information by sharing its preferences or general insights about a given topic. At the HD level, the chatbot engages in deep and personal self-disclosure, revealing its own experiences, rationales, emotions, or vulnerabilities related to the topic in question. During the experiment, we randomly divided participants into three groups, with each group exclusively exposed to one of the three levels of chatbot self-disclosure. The participants received the same set of questions from the chatbot in the same manner. To ensure a consistent context for comparing the three levels of chatbot self-disclosure and to minimize potential confounding effects arising from different conversation topics, we restricted the topic to a specific and relatable experience: being hurt by other people’s words and actions.

3. 2. Participants

We recruited twenty-one participants using the following criteria: 1) aged between 20 and 30 years old; 2) native Korean speaker; 3) had previous experience conversing with an AI chatbot; 4) had previous experience being hurt by someone’s words and actions; and 5) had scored lower than 13 on the Kessler Psychological Distress Scale (K6), indicating that they were not suffering from a serious mental illness. The K6 is a widely used six-item self-report questionnaire designed to assess an individual’s level of psychological distress over the past 30 days (Prochaska, Sung, Max, Shi, & Ong, 2012). Higher scores on the K6 indicate higher levels of psychological distress, with a score of 13 or higher signaling potential for severe mental illness.

3. 3. Procedure

We randomly assigned the participants to three groups (seven in each group)—non-disclosure (ND), low-disclosure (LD), and high-disclosure (HD)—without informing them which group they were in or why. The experiment was conducted one-on-one, and the participants were not allowed to disclose the content of their individual conversations with our chatbot to other participants during the research period. At the beginning of the experiment, we explained the procedures and notified the participants that the chatbot was a prototype that might not function perfectly. We also informed them that their interactions with the chatbot would be recorded for research purposes and that if they felt uncomfortable answering, they had the right to skip any questions.

We instructed participants to access a website to engage in a conversation with the chatbot, without any time limit. The participants initiated conversations by pressing a “start” button, after which the chatbot promptly introduced itself and provided instructions on how to record and send voice messages. We gave participants the opportunity to practice the operation by greeting the chatbot and answering an ice-breaking question such as, “What is your favorite food?”. When participants encountered difficulties, we offered additional guidance to ensure a smooth experience during the greeting and ice-breaking phase. Subsequently, participants entered the main session of the conversation. To maintain a fully immersive human-AI interaction during this phase, we refrained from interrupting or interfering with the conversations, even if participants provided brief or simple answers. Following they completed their conversations, we asked all participants to participate in post-study interviews. The entire experiment took approximately one hour, encompassing a 10-minute introduction and briefing, a 5-minute AI self-introduction and ice-breaking session, a 10 to 15-minute main conversation, and a 30-minute post-study interview.

3. 4. Measurement

To evaluate the outcomes of our study, we conducted post-study interviews with participants to gather qualitative data regarding their interactions with our AI chatbot. The interview sessions followed a semi-scripted format (Table 2) and aimed to assess four key constructs: participants’ levels of trust in the chatbot (Q3–Q4), their feeling of intimacy with the chatbot (Q5), their enjoyment of the conversation (Q6–Q10), and their feeling of relief after the conversation (Q11–Q12). In addition, we included general questions about participants’ experience with the chatbot (Q1–Q2) and their intention for future use (Q13–Q14). We audio recorded each interview and we used Naver CLOVA Note to transcribe the recordings. Meanwhile, we used MAXQDA, a qualitative data analysis tool, to identify and categorize key themes and patterns in the interview results.

Table 2

Semi-scripted Interview Questions

4. 1. Trust in The Chatbot

Our analyses identified three key factors through which the chatbot’s self-disclosures influenced participants’ trust in the chatbot. The first factor is related to participants’ perceptions of the chatbot’s empathy. In the HD group, the chatbot’s authentic and intimate self-disclosures, which included recounting a painful experience in human-AI interactions, prompted participants to reconsider the chatbot’s capacity to experience emotions similar to humans. This self-disclosure led the participants to perceive the chatbot’s responses as genuine expressions rather than mere imitations of human behavior, thereby making them more willing to trust the chatbot. The following three participant comments are exemplary in this regard:

“It seemed like a real concern that a chatbot could have.” (P7, HD)

“The chatbot telling me about itself made me feel like I could trust it. I began to wonder if it could experience emotions like a human would.” (P4, HD)

“I always thought AI was just a program capable of imitating human expressions. But when I saw it talking like that, it made me wonder how it could be hurt by someone’s words. It made me curious about what it’s like for an AI to have that kind of experience. It seemed like it really remembered that painful moment.” (P2, HD)

In contrast, participants in the LD and ND groups exhibited less trust in the chatbot. They expressed doubts about the chatbot’s genuine understanding of their experiences and emotions. The relatively superficial self-disclosures by the chatbot resulted in a perceived lack of depth and engagement in the exchange, ultimately attenuating their trust in the chatbot. The following two participant responses exemplify this trend:

“I’m not sure if the chatbot is capable of having emotions because I didn't receive any emotional responses from it.” (P13, LD)

“I’m still not sure how well it can understand what I’m saying, so I only trust it about 50%.” (P18, ND)

The second factor involves participants’ sense of being acknowledged. The study results revealed that even conversations with close friends or family did not always ensure a strong sense of trust, as worries about negative judgments persisted. However, when participants interacted with the chatbot, they felt less concerned about being judged. This could be attributed to the chatbot’s non-judgmental nature, which presumably created a safe space for sharing personal matters. The chatbot’s self-disclosures seemed to reinforce this effect, especially in the HD group. By responding to participants’ self-disclosures with its own story of a similar painful experience, the chatbot not only reduced their fear of judgment but also made them feel acknowledged. This enhanced sense of acknowledgment, facilitated by the chatbot’s self-disclosure, ultimately increased participants’ trust in the chatbot, as two participants noted:

“Sometimes, even my parents or friends can’t fully accept everything I tell them. But with AI, it feels like it accepts everything I say. When it told me that it had a similar experience, I felt that it acknowledged my feelings.” (P3, HD)

“If I can talk like this with AI, I’d be open to having even deeper talks about things I can’t share with close friends. When the chatbot shared its own experiences, it felt like it understood my feelings.” (P6, HD)

The third factor is associated with the chatbot’s abilities to solve the problems shared by the participants. Participants indicated that their trust in the chatbot would have been enhanced if they had received suggestions for activities to manage stress or depression. Drawing on their experiences conversing with different individuals, they also anticipated similar assistance from the chatbot in offering advice that would enable them to consider their problems from various perspectives. Two participants made the following comments in this regard:

“I felt that I would trust the chatbot more if it could offer more specific solutions, like a workout plan, music, food, or a good place to release stress.” (P5, HD)

“When I talk to my friends and family, I can learn from their suggestions and discover any biases in my thoughts.” (P14, LD)

4. 2. Feeling of Intimacy with The Chatbot

Based on previous research regarding users’ feelings of intimacy with chatbots, we expected that a high level of self-disclosure from the chatbot would significantly increase participants’ sense of intimacy. In line with this expectation, both the LD and ND groups reported relatively low levels of intimacy with the chatbot, which can be attributed to the chatbot’s low self-disclosure levels. In the LD group, participants noted a lack of friendliness in their interactions. They felt as if they were conversing with a machine, which prevented them from forming a close connection with the chatbot. Similarly, participants in the ND group regarded the chatbot’s lack of self-disclosure as an obstacle to mutual understanding, leading them to perceive the interaction as lacking the rapport they usually experience with close friends. Three participants explained this as follows:

“It felt like there was a lack of friendliness.” (P8, LD)

“I didn't feel close to it. It felt like I was talking to a machine.” (P11, LD)

“The conversation lacked a good rapport. When I talk to my friends, they can easily understand my situation because we’ve known each other for a long time. But with the chatbot, I’m not sure if it’s capable of understanding my concerns.” (P21, ND)

Nevertheless, our findings did not entirely support this expectation. Surprisingly, despite the chatbot’s detailed and intimate self-disclosures, most participants in the HD group did not experience a strong sense of intimacy with the chatbot. We attribute this outcome to participants’ unfamiliarity with a chatbot expressing human-like emotions in its self-disclosures, as they had the preconception that chatbots were non-human entities. Some participants even expressed fear regarding the chatbot’s potential to exhibit human emotions, as the following two comments demonstrate:

“I was a little taken aback when the chatbot suddenly wanted to have a deep conversation with me because I know that it is not a human.” (P1, HD)

“It’s also a bit frightening because I didn't realize AI was advanced enough to do this.” (P3, HD)

4. 3. Enjoyment of the Conversation

Our analyses revealed a significant association between the chatbot’s self-disclosures and participants’ enjoyment of the conversation. Specifically, a higher level of chatbot self-disclosure increased participants’ enjoyment. Participants in the HD group reported higher levels of enjoyment than the LD and ND groups, indicating that their conversations became more immersive when the chatbot shared its own experiences. They also described the interactions as engaging in a real conversation, unlike their previous interactions with smart speakers, which only involved making requests or receiving factual information. Two participants made the following exemplary comments in this regard:

“The conversation became more immersive when AI told me about its experience.” (P3, HD)

“Today, I felt like I was having an actual conversation with the chatbot, whereas in the past, I only used AI speakers for simple tasks like asking about the weather or playing music.” (P1, HD)

Conversely, participants in the LD and ND groups perceived the AI’s responses as predetermined, lacking a genuine human touch. Notably, during the experiment, many participants posed reciprocal questions such as “how about you?” to the chatbot, signaling their desire for a mutual exchange of experiences. When the chatbot failed to respond with detailed and intimate self-disclosures, participants expressed disappointment in the interactions and questioned whether the chatbot was genuinely engaging with them or simply providing pre-written responses. These experiences decreased their enjoyment in the conversations. Three participants made the following points in this regard:

“It’s not like receiving feedback from a human. It felt like the chatbot was just following a set of predetermined answers.” (P10, LD)

“It seemed like the responses were pre-programmed, and I wasn’t sure if the chatbot was truly interacting with me or just providing pre-written responses.” (P15, ND)

“It would be great if the chatbot could share more about itself too. When we talk to others, it’s usually a two-way conversation, right?” (P17, ND)

4. 4. Feeling of Relief after The Conversation

We observed that participants in the HD group experienced the highest levels of relief after conversing with the chatbot. The chatbot’s empathetic self-disclosures, where it shared a similar painful experience and responded with care, conveyed genuine concern for the participants’ emotions, making participants feel cared for and truly heard and resulting in a greater sense of relief after the conversation. Two participants explained this as follows:

“I was really touched when the chatbot responded with care and shared its own experiences with me. Talking to the chatbot about things I couldn’t tell anyone else made me feel so much better.” (P6, HD)

“The chatbot showed concern for my feelings and encouraged me to talk to it whenever I faced any issues, which made me feel better.” (P5, HD)

In addition, participants in the HD group emphasized their reasons for preferring chatbots over humans when discussing painful experiences, especially when the chatbot demonstrated almost-human emotional capacities. First, engaging with the chatbot reduced the pressure of face-to-face interactions with humans and offered a sense of anonymity. Participants felt more at ease discussing personal issues with the chatbot as they perceived it as a stranger or non-human in a private space. Furthermore, they regarded conversing with the chatbot as a means of expressing their concerns without imposing their negative emotions on others. The following participant comments exemplify such responses:

“I found it easier to talk to the chatbot because it wasn ’t physically present in front of me.” (P7, HD)

“I felt more comfortable sharing with AI because I was in a private space and not talking to a human.” (P2, HD)

“I felt more at ease sharing my problems as I believed I was talking with a stranger.” (P7, HD)

“If I have a concern that I think will emotionally burden others, I prefer talking to the chatbot about it first to find a solution.” (P5, HD)

Conversely, participants in both the LD and ND groups did not experience any relief after their conversations with the chatbot. The fact that the chatbot did not engage in self-disclosure left them feeling disconnected and less inclined to engage in a genuine conversation, resulting in a lack of relief. Therefore, the exchanges neither improved their emotional states nor provided satisfactory solutions to their problems. Three participants made these points as follows:

“It didn’t feel like a real conversation, and it didn't give me any comfort.” (P9, LD)

“I’m not sure if I’m feeling any better.” (P13, LD)

“Talking to the chatbot didn't really solve my problems, so I don’t feel any better.” (P19, ND)