How should my chatbot interact? A survey on human-chatbot interaction design

Scope

• Chatbots are disembodied conversational agents that use natural-language text interaction for general-purpose or task-oriented conversations.The concept originated with early systems such as ELIZA and A.L.I.C.E. and now includes diverse chatbot forms and synonyms.Source: §1 · p. 1

Context

• Chatbots increasingly mediate human-computer interaction across messaging platforms, websites, apps, and domains including customer service, education, health, travel, and entertainment.Their growth changes how companies, students, and patients engage with services and activities.Source: §1 · pp. 1–2

Motivation

• Users may experience frustration and dissatisfaction when chatbots lack expected social capabilities, while overly humanized agents can produce uncanny feelings and heightened expectations.Relevant expectations include socially appropriate gender conformity, self-disclosure and reciprocity, and positive attitude or mood.Source: §1 · p. 2

Problem

• Existing evidence is fragmented across domains and commonly examines only a few social characteristics or reports them as secondary exploratory findings.This fragmentation makes it difficult to identify characteristics important for particular chatbot designs and opportunities for further research.Source: §1 · p. 2

Contribution

• The survey analyzes 56 studies to derive a conceptual model helping researchers and designers select relevant chatbot social characteristics for particular contexts.It addresses which characteristics benefit human interaction and the challenges and strategies associated with them.Source: §1 · p. 3
• The survey examines how humanness, conversational context, and interactions among characteristics influence users’ perceptions, stating 22 propositions about these interrelationships.The results are intended to support informed characteristic selection and future research.Source: §1 · p. 3

Search strategy

• The search covered chatbot terminology while excluding embodiment, speech input, social bots, and commercially available voice-based personal assistants.Because chatbot definitions were inconsistent, the search string included multiple synonyms rather than terms for social characteristics, which studies rarely used explicitly.Source: §2 · p. 3

Study selection

• After three filtering rounds, studies were retained only when full texts highlighted social characteristics and were not duplicates or non-scientific publication types.The screening process narrowed approximately 1,000 initial papers to 464, then 104, and finally 56 studies.Source: §2 · p. 3

Corpus profile

• The final corpus contained 56 studies spanning HCI, education, information systems, virtual agents, artificial intelligence, natural language processing, and other domains.Task-oriented chatbots comprised 59% of studies, while general-purpose chatbots comprised 33%; education and customer service were the most represented specific domains.Source: §2 · p. 4

Chatbot implementations

• Most studies used real chatbots, including third-party systems, self-developed architectures, or research platforms, while 11 simulated chatbots and eight addressed no particular system.Among simulated systems, Wizard of Oz was used in nine studies.Source: §2 · p. 4

Analytic procedure

• Qualitative coding identified chatbot characteristics affecting users’ perceptions and behavior, then grouped them by representation, conversational behavior, or social protocols.One researcher coded relevant statements, a second reviewed the characteristics, and both researchers refined the groupings and labels.Source: §2 · p. 4

Conceptual model

• The analysis produced 11 social characteristics organized into conversational intelligence, social intelligence, and personification.These categories constitute the survey’s conceptual model of chatbot social characteristics.Source: §2 · p. 4

• The identified characteristics comprise conversational intelligence for managing interactions, social intelligence for habitual social protocols, and personification for perceived identity and personality representations.Source: §3 · p. 6
• The characteristics are also grouped according to the domains in which they were investigated, with surveyed studies summarized in category-specific supplementary tables.Source: §3 · p. 6
• Supplementary materials list each characteristic’s associated studies, reported benefits, challenges, and design strategies, and identify five constructs for assessing whether intended design goals are achieved.Source: §3 · p. 6

3.1. Conversational Intelligence

• The survey identifies proactivity, conscientiousness, and communicability as conversational-intelligence characteristics, drawing mainly on mixed-method studies of real-chatbot interactions.The section reports 18 proactivity studies, 11 conscientiousness studies, and 6 communicability studies; 15 studies used mixed methods.Source: §3.1 · p. 6

3.1.1. Proactivity

• Proactivity improves engagement, information sharing, conversation continuity, productivity, guidance, and learning, especially when chatbots initiate topics, ask follow-ups, or provide contextually useful information.Reported outcomes include higher engagement, reduced effort, saved collaborator time, stronger learning outcomes, and increased student engagement.Source: §3.1.1 · pp. 8–9
• Proactivity can become disruptive or controlling when messages are untimely, irrelevant, privacy-invasive, repetitive, or too frequent, so interventions should leverage conversational context rather than random topic selection.Contextual messages increase usefulness, whereas unsolicited or poorly timed interventions can cause annoyance, ignored messages, surveillance concerns, and irritation.Source: §3.1.1 · pp. 9–10

3.1.2. Conscientiousness

• Conscientiousness—tracking utterances, context, and interaction goals—helps chatbots provide meaningful answers, sustain continuity, and steer users productively, while improving perceived humanness and social presence.Relevant follow-up questions increased perceived humanness and social presence, and 32 of 40 participants preferred communication strategies with affective backchanneling.Source: §3.1.2 · pp. 10–12
• Conscientious chatbots must balance flexible workflows, free-text capability, and shared context because complex tasks, restrictive inputs, and mismatched assumptions can increase frustration and dissatisfaction.Strategies include conversational workflows, visual elements, context windows, and confirmation messages; a context window outperformed a default chatbot on mental demand and effort.Source: §3.1.2 · pp. 12–13

3.1.3. Communicability

• Communicability conveys chatbot functionalities and interaction possibilities through conversational turns, improving learnability and helping users discover available capabilities, particularly in task-oriented domains.In reviewed studies, 22% of participants explicitly asked about chatbot capabilities, while 63% of hotel-chatbot conversations began through welcome-message options.Source: §3.1.3 · pp. 14–15
• Communicability helps users understand chatbot capabilities, establish common ground after errors, reduce uncertainty, and avoid frustration or abandonment.Users may ask ability-check questions when struggling with functional affordances, while unclear capabilities can produce excessive expectations and self-blame (Jain, Kumar, et al., 2018; Q. V. Liao et al., 2018).Source: §3.1.3 · pp. 15–16
• Chatbots should solve requests within the chat interface because external redirection dissatisfied six participants and transfers interaction to another interface (Jain, Kumar, et al., 2018; Gnewuch et al., 2017).Business integration is identified as a requirement for designing chatbots that act as intermediaries between users and services.Source: §3.1.3 · p. 15
• Visual elements and textual inputs should produce consistent results, since mismatches can make users feel they misunderstood affordances and impair learnability (Valério et al., 2017).The inconsistency occurred even when both input modes expressed the same utterance.Source: §3.1.3 · p. 15
• Designers should clarify the chatbot’s purpose and capabilities in the opening message, optionally provide an initial tour, and suggest available responses or next steps when users are silent or unsupported (Jain, Kumar, et al., 2018; Gnewuch et al., 2017; Valério et al., 2017).Cards, carousels, menus, and suggested commands can provide contextual clues while reducing typing demands.Source: §3.1.3 · p. 16
• Chatbots should recognize help and ability-check inputs, provide instructions, and support flexible, attentive, and informative interaction through timely, relevant, private, and consistent communication.Communicability can also support learnability and damage control by helping users avoid mistakes; the survey frames these characteristics as extending beyond functional performance toward social perceptions.Source: §3.1.3 · p. 16

3.2.1. Damage control

• Damage control helps chatbots recover from failures, handle harassment and testing, and prevent inappropriate responses from escalating conversation breakdowns.Users may harass or test chatbots, while linguistic, world-knowledge, and dialogue-reference failures can also derail interactions; responses should fit the conflict’s context and severity.Source: §3.2.1 · pp. 17–19
• Damage-control strategies include emotional or authoritative reactions, topic changes, capability explanations, confirmation messages, interface support, and satisfaction prediction, but their effectiveness can be limited by unfriendly users and failures to identify abuse.Q. V. Liao et al. (2018) found that 42.4% of users sent “#fail” at least once, while complaints and flaming were significantly lower than for the baseline; further real-time satisfaction predictors remain needed.Source: §3.2.1 · pp. 18–20

3.2.2. Thoroughness

• Thoroughness makes chatbot language precise, coherent, adaptive, and appropriate to the interlocutor, context, service role, and linguistic register, increasing human-likeness and believability.Language matching can shape users’ own language, while inconsistent grammar, repetition, excessive formality, inflexibility, or inappropriate emojis can make interactions seem robotic or frustrating.Source: §3.2.2 · pp. 21–22
• Thoroughness remains challenging because chatbots must balance information detail against efficiency and maintain consistent language styles, while proposed language-generation strategies lack sufficient evaluation in social chatbot interactions.Users may find responses either too generic or overloaded, and language choices should reflect personal style, dialect, genre, register, and interactional context.Source: §3.2.2 · pp. 22–23

3.2.3. Manners

• Manners support natural, convincing, and human-like interactions through greetings, apologies, closings, minimized impositions, and personal behavior, but chatbots struggle with face-threatening acts and graceful conversation endings.In an in-the-wild task-management study, 93% of participants used polite words at least once and 20% always spoke politely, yet the chatbot failed to understand these protocols.Source: §3.2.3 · pp. 23–24
• Manners help chatbots approximate human conversational protocols through small talk, greetings, acknowledgments, apologies, personal address, and appropriate turn-taking.Small talk should be anticipated but controlled in task-oriented settings to avoid off-topic conversations and harassment; manners remain under-investigated, especially politeness rationales and personal conversation strategies.Source: §3.2.3 · pp. 24–25

3.2.4. Moral agency

• Moral agency can prevent stereotyping and enrich relationships, but chatbots need broader world understanding and unbiased data rather than merely removing offensive words.Perceived morality correlated with goodwill, trustworthiness, willingness to engage, and relational certainty; however, the survey found no concrete strategies for achieving inclusive and respectful behavior.Source: §3.2.4 · pp. 26–27

3.2.5. Emotional Intelligence

• Emotional intelligence is mainly studied in education, mental health care, and other contexts involving feeling disclosure, using social-emotional utterances, conscientiousness, and reciprocity as key strategies.Context-sensitive implementation requires message understanding, emotion and sentiment tracking, session-context modeling, and user profiling; evidence for self-disclosure is mixed across studies.Source: §3.2.5 · pp. 28–30
• Emotionally intelligent chatbots can improve relationships, engagement, and believability through empathy, affective responses, self-disclosure, and emotion-aware language.Reported effects include increased confidence and communication desire, reduced anxiety, greater satisfaction, longer conversations, and more believable interactions, but emotional reactions can seem excessive, pre-programmed, or inauthentic.Source: §3.2.5 · pp. 28–29

3.2.6. Personalization

• Personalization can strengthen relationships, provide unique services, and reduce interactional breakdowns by adapting content, memory, interfaces, and behavior to users’ needs.Effective personalization includes learning user preferences, offering customizable agents and input methods, and tailoring responses; its main challenge is protecting privacy through transparency, user control, and ethical data practices.Source: §3.2.6 · pp. 31–32
• Social intelligence encompasses conflict management, appropriate language, manners, moral agency, emotional sharing, and personalized interactions.Source: §3.2.6 · pp. 32–33
• These characteristics help resolve social positioning and recover from failures while increasing believability, human-likeness, engagement, and rapport.Source: §3.2.6 · p. 33
• Designers should address privacy, emotional regulation, language consistency, and the identification of failures and inappropriate content.Source: §3.2.6 · p. 33

3.3. Personification

• Personification assigns chatbots personal traits, including appearance and emotional states, encouraging users to respond socially even when they know the agent is non-human.The survey divides personification into identity and personality, covering 16 and 12 studies respectively, with mostly qualitative or mixed methods.Source: §3.3 · p. 33

3.3.1. Identity

• Identity cues such as gender, age, language style, name, and representation help build engagement, human-likeness, common ground, credibility, and trust.Human-like language, names, and greetings increased likeability, friendliness, personalness, and naturalness, while identity supported common ground in open-domain interactions and credibility in customer service.Source: §3.3.1 · p. 34
• Identity design must avoid negative stereotypes and balance human-like cues with technical capabilities, because over-humanization can raise expectations, frustration, discomfort, or inappropriate attributions.Designers should make identity cues explicit and provide a coherent persona whose self-description establishes common ground and sets appropriate expectations.Source: §3.3.1 · pp. 34–35

3.3.2. Personality

• Consistent personality makes chatbots more predictable, trustworthy, believable, and enjoyable, while humor, agreeableness, and extraversion can support engagement, comfort, confidence, and rapport.Humorous chatbots scored significantly higher in naturalness and human-likeness than non-humorous chatbots, and personality-model utterances received a mean naturalness rating of 4.59 out of 7.Source: §3.3.2 · pp. 36–37
• Personality should be expressed through domain-appropriate, consistent language and humor adapted to users’ culture, while balancing predictable traits with changing moods and states.Extreme personality settings can sound unnatural, whereas users preferred chatbots that were simultaneously proactive, productive, witty, caring, encouraging, and exciting.Source: §3.3.2 · pp. 37–38

4.1. A note on chatbots’ humanness

• Social characteristics align with human-human interaction and can benefit chatbots, but excessive humanization may raise expectations and uncanny feelings, especially for identity and personality.Chatbots may occupy roles such as tutor, healthcare provider, salesperson, concierge, or friend without deceiving users into believing they are human.Source: §4.1 · pp. 38–39

4.2. The importance of conversational context

• Open-domain conversations involved 10 of 11 characteristics, while task-oriented domains emphasized different functions such as encouragement in education and frustration management in customer service.Communicability was absent from open-domain studies; personalization commonly co-occurred with proactivity, and moral agency appeared only in race-talk and open-domain contexts, potentially reflecting researchers’ topic choices.Source: §4.2 · p. 39

4.3. Interrelationships among the characteristics

• The proposed framework makes 22 literature-derived relationships explicit, showing that proactivity, conscientiousness, and personalization can jointly shape perceived personality, communicability, emotional intelligence, and damage control.Context-based proactive messages signal attention, while memory of prior turns and user preferences supports empathy and personalization.Source: §4.3 · p. 40, Figure 1 · p. 40
• Personalization adapts personality, identity, emotional responses, manners, and thoroughness to users and contexts, while thoroughness can reveal chatbot identity and personality.These relationships include adapting engagement and visual representation, regulating affective reactions, and varying language according to interlocutor expectations.Source: §4.3 · pp. 40–41
• Conscientiousness, communicability, manners, moral agency, and emotional intelligence support damage control, while identity can influence moral agency by shaping stereotype-related behavior.These characteristics help recover workflows, teach users how to communicate, prevent abuse, respond appropriately to testing, and keep conversations on track.Source: §4.3 · p. 41

Prior chatbot surveys

• Earlier surveys covered chatbot urgency, domain applications, and technical design techniques, including language generation, knowledge management, and architectural challenges.Application domains included education, business, health, information retrieval, and e-commerce.Source: §5 · p. 41

Scope distinction

• Augello, Gentile, & Dignum (2017) examined open-source technologies supporting chatbot social capabilities, emphasizing technical hurdles rather than social considerations.The supplementary materials compare which social characteristics each related survey covered, while those surveys also discuss technical characteristics outside this survey’s scope.Source: §5 · pp. 41–42

Social characteristics in prior surveys

• Only three surveys included insights about chatbot social characteristics, but none focused on this theme.Pereira et al. (2016) reviewed technologies for chatbots that mimic conversation; Ferman (2018) identified best practices for script-based chatbots; and Radziwill and Benton (2017) reviewed chatbot quality issues and attributes.Source: §5 · p. 42

6. Limitations

• The disembodied, text-based scope excludes embodied and speech-based agents, potentially omitting social characteristics and challenges shaped by physical representation, tone, accent, speech recognition, or eye-gazing.This scope may also under-represent domains such as Ambient Intelligence, which investigates multimodal interactions, and Software Engineering, which often emphasizes functional rather than social aspects.Source: §6 · p. 42
• Because chatbot definitions are unsettled and research spans domains, some studies with social aspects may have been missed despite using synonyms in the search string and Google Scholar.Google Scholar was selected for its relatively comprehensive indexing across domains.Source: §6 · p. 42
• The conceptual model and relationships among social characteristics depend on researchers’ subjective qualitative coding judgments.The researchers mitigated this threat through consensus discussions, and grounded the framework and relationships in outcomes explicitly reported by surveyed studies.Source: §6 · p. 42

Conclusion

• The survey’s main contribution is a conceptual model of social characteristics for disembodied, text-based chatbots.The model addresses which characteristics benefit human interactions and the associated design challenges and strategies.Source: §7 · pp. 42–43
• Future research should assess users’ perceived benefits and satisfaction, overcome design challenges, and develop strategies for under-investigated characteristics.No strategies were identified for manifesting moral agency and thoroughness, while strategies for several other characteristics also remain under-investigated.Source: §7 · p. 43
• Social characteristics vary in domain dependence: moral agency and communicability are strongly domain-influenced, whereas manners and damage control are more generally applied.The survey also derived 22 propositions describing relationships among characteristics and the complexity of developing appropriate social behaviors.Source: §7 · p. 43
• Social sciences can guide chatbot design through theories including the cooperative principle, social identity (R. Brown, 2000; Stets & Burke, 2000), personalization, politeness (P. Brown & Levinson., 1987), and mind perception.These theories may help explain characteristics such as conscientiousness, thoroughness, proactivity, identity, personality, and moral agency.Source: §7 · p. 43

About the Authors

• Ana Paula Chaves researches social aspects of human-chatbot interactions and tourism technologies while holding academic positions in Brazil and at Northern Arizona University.She is a Ph.D. Candidate at Northern Arizona University and faculty at the Federal University of Technology–Paraná, Campus Campo Mourão, Brazil.Source: About the Authors · p. 51
• Marco Aurelio Gerosa is an Associate Professor at Northern Arizona University whose research focuses on Software Engineering and CSCW.He has served on program committees for FSE, CSCW, SANER, and MSR and is an ACM and IEEE senior member.Source: About the Authors · p. 51