Source-linked AI summary
What Makes a Good Conversation? Challenges in Designing Truly Conversational Agents
Leigh Clark, Nadia Pantidi, Orla Cooney, Philip Doyle, Diego Garaialde, Justin Edwards, Brendan Spillane, Christine Murad, Cosmin Munteanu, Vincent Wade, Benjamin R. Cowan
TL;DR
Conversational agents often provide constrained, task-oriented interaction, while human conversation supports both practical goals and social relationships. Through semi-structured interviews, the paper identifies valued conversational characteristics and compares how participants understand them in human and agent contexts. Participants treated agent conversation as almost purely transactional, questioned relationship-building with agents, and highlighted context and relationship dynamics as central design challenges.
Problem
Current agents emulate human conversational features without adequately addressing the subjective qualities users value or how those qualities should vary in agent interactions.
Method
The paper uses semi-structured interviews to compare people’s views of conversational characteristics in human-human and human-agent contexts.
Results
Participants distinguished social and transactional human conversation but described agent conversation almost purely in transactional terms, framing common ground as one-way personalisation.
Takeaways & Limitations
Conversational-agent design must account for context, interaction purpose, and relationship dynamics rather than assume human-like conversational structures reproduce human social outcomes.
Takeaways & Limitations
Participants did not interact with conversational agents during the study, and most were intermediate, advanced, or expert technology users.
Abstract
from arXiv · showhide
Conversational agents promise conversational interaction but fail to deliver. Efforts often emulate functional rules from human speech, without considering key characteristics that conversation must encapsulate. Given its potential in supporting long-term human-agent relationships, it is paramount that HCI focuses efforts on delivering this promise. We aim to understand what people value in conversation and how this should manifest in agents. Findings from a series of semi-structured interviews show people make a clear dichotomy between social and functional roles of conversation, emphasising the long-term dynamics of bond and trust along with the importance of context and relationship stage in the types of conversations they have. People fundamentally questioned the need for bond and common ground in agent communication, shifting to more utilitarian definitions of conversational qualities. Drawing on these findings we discuss key challenges for conversational agent design, most notably the need to redefine the design parameters for conversational agent interaction.
1 Introduction
The paper examines why current conversational agents fall short of human conversation and investigates which conversational characteristics people value in human and agent interactions. Interviews identify a gap between emulating conversational behavior and addressing users’ subjective expectations.
- Current agent interactions remain constrained and task oriented despite the growing importance of human-agent conversation.
- Developers have emphasized human-like features such as social talk and humour, but this may overlook users’ subjective conversational expectations.
- The study uses interviews to identify important conversational characteristics and how they vary between human and artificial agents.
- Participants distinguished social and transactional conversation, while agent conversations were discussed almost universally in transactional terms.
2 Related Work
Prior work distinguishes social conversation from task-based transaction and shows that many current assistants remain brief, functional, and less adaptive than human dialogue. The paper therefore examines how conversational characteristics differ across human and agent contexts.
- Human conversation is broadly classified as transactional, pursuing practical goals, or social, building and maintaining relationships.
- Conversation involves opening an interactional channel, mutual engagement, co-constructing meaning, and converging on agreement.
- Current IPAs mainly support brief, clearly delineated tasks such as information seeking, reminders, and device control.
- IPA exchanges often use question-answer or instruction-confirmation formats and rarely reach elaborative, contextual social talk.
- The study contrasts established human-human conversational properties with less-understood perceptions of conversations with machines.
3 Method
The study recruited university-community participants and used semi-structured interviews to explore conversational characteristics across human and agent contexts. Interviews were transcribed and analysed through inductive thematic analysis.
- Seventeen university-community participants were recruited until saturation, with a mean age of 31.1 years.
- Most participants had used voice assistants, although 58.8% of that group reported using them very infrequently.
- Interviews averaged 40 minutes and covered conversation characteristics, purposes, experiences, and attitudes toward agents.
- Two researchers independently coded transcripts before researchers reviewed and grouped themes through systematic inductive thematic analysis.
4 Findings
Participants described conversation as serving both social and transactional purposes, with valued qualities depending on conversational purpose and partner. They associated human conversation with relationship development, but framed agent interactions more functionally.
- Social and transactional purposes: Conversation served social purposes such as establishing, maintaining, and deepening social bonds.
- Social and transactional purposes: Transactional conversations pursued clear short-term objectives and often ended once the goal was achieved.
- Conversational attributes: Participants valued common ground, including mutual understanding of speakers’ intent and meaning.
- Conversational attributes: Trustworthiness supported common ground and sustained long-term relationships, while also enabling more personal conversation.
- Conversational attributes: Active listening involved attention, engagement, understanding, and reciprocal participation in two-way dialogue.
- Conversational attributes: Humour was commonly valued as a conversational quality that could add substance to interaction.
Humour
Conversation varies with relationship context: social interaction can mix humour with substantive or serious discussion, while interactions with strangers remain more limited and functional. Humour helps sustain human conversation but is treated as a novelty rather than a necessity with agents.
- Humour: Effective humour was expected to have substance and relevance, sometimes softening serious intentions or delivering substantive messages.
- Relationship context: Conversational needs and topics changed with relationship type and stage.Close friends relied on trust and shared history, while strangers and acquaintances typically discussed superficial or functional matters.
- Conversing with friends: With very close friends, silence could signal comfort rather than a need for constant conversation.
- Strangers and acquaintances: Small talk and transactional dialogue helped strangers feel comfortable and fulfilled social expectations, especially during silence.
- Transition toward friendship: Conversation helped people transition toward friendship by sharing vulnerabilities, discovering common interests, and building trust through repeated experiences.
Transition Towards Friendship
Participants saw conversation as a pathway toward friendship, but agent interactions were perceived as fundamentally different. Human conversational qualities such as common ground, trust, and understanding were translated into more functional agent requirements.
- Transition Towards Friendship: Conversation helped people gauge connections and develop friendship through vulnerability, mutual understanding, and trust.
- Transition Towards Friendship: Common ground developed through shared interests, traits, and repeated experiences.
- Agent interaction: Participants perceived a high barrier to achieving the social and emotional connections associated with human conversation with agents.
- Agent interaction: Participants questioned whether human conversation could or should be emulated, instead calling for different design parameters for machine conversation.
- Agent interaction: In agent dialogue, common ground was not viewed as co-constructed; instead, the agent was expected to lead personalization.
- Agent interaction: Trustworthiness in agent conversations emphasized security, privacy, transparency, efficiency, and reliability rather than emotional trust.
Functional Trustworthiness and Privacy
Participants defined trustworthiness and listenership functionally in agent conversations, prioritizing data security, privacy, transparency, reliability, and accurate speech recognition. Humour could make agents more interesting, but was rarely considered necessary.
- Functional Trustworthiness and Privacy: Machine trustworthiness centered on security features and control over recorded data, its use, and access.
- Functional Trustworthiness and Privacy: Agent trustworthiness lacked the emotional emphasis associated with human trust but could support frequent long-term use through efficiency, reliability, and security.
- Functional listening: Participants expected agents to understand speech clearly and quickly without requiring repetition.Many comments focused on speech-recognition performance.
- Humour: Agents’ humour was viewed as a novelty that could make interactions more interesting, rather than as an organic conversational process.
- Humour: Humour was rarely described as necessary for agents in the way participants considered it relevant to human conversation.
- Functional interaction: Current human-agent conversations primarily served transactional purposes, resembling interactions with strangers or casual acquaintances.
"I mean I don’t enjoy communicating with machines...when
Participants generally rejected friendship with machines because agent interaction felt unlike human conversation, while recognizing possible value in companionship and practical support. This creates tensions around effort, commercial incentives, agency, and the boundary between relationship and tool.
- Relationship Tensions: Participants often considered friendship with machines unnatural or incompatible with how they converse with human friends.
- Relationship Tensions: Building a relationship with an agent was seen to require substantial time and effort and might change the nature of friendship.
- Relationship Tensions: A master-servant relationship conflicted with friendship because users could order agents while friendship might imply an agent’s right to refuse.
- Relationship Tensions: Commercial incentives created tension because chatbots were associated with cheaper, more flexible customer care rather than conversation itself.
- Potential value: Participants saw potential value for isolated people, including older adults or those experiencing mental-health difficulties, through companionship.
- Potential value: Participants primarily proposed functional applications such as appliance control, health monitoring, scheduling, and administrative support.
5 Discussion
Interviews distinguished social and transactional purposes in conversation, but participants primarily understood agent conversation through transactional goals and functional performance. They also questioned whether agents could or should support friendship-like bonds.
- Mutual understanding, trustworthiness, active listenership, and humour mattered in both contexts but were operationalized functionally for agents.
- In agent conversations, personalisation created an illusion of common ground without collaborative negotiation and updating during dialogue.
- Participants questioned befriending agents, citing master-servant dynamics, monetary incentives, and status asymmetry between users and agents.
Current Perceptions as a Barrier to Conversational Agent Interaction
Participants viewed agents as user-controlled tools rather than companions or social equals, shaping conversation toward functional and transactional expectations. This perceived mismatch may restrict the conversations users consider appropriate or possible.
- Agents were perceived as user-controlled tools rather than potential companions or social equals.
- Functional ability and status asymmetry may restrict the types of conversations users perceive as appropriate or possible with agents.
- Social interaction was largely absent from participants’ perceptions of what conversational agents could and should perform.
Reframing the Concept of Conversation in Agent Interaction
The findings support reframing agent interaction as a distinct conversational genre with functional norms rather than simply simulating human conversation. Appropriate norms depend on context and may change as agents are used for longer-term social needs.
- Reframing the Concept of Conversation in Agent Interaction: Agent interaction should be treated as a new conversational genre with its own rules, norms, and expectations.
- Reframing the Concept of Conversation in Agent Interaction: These norms may be more functional and utilitarian, with less emphasis on relational growth or emotional outcomes than human conversation.
- Reframing the Concept of Conversation in Agent Interaction: Conversational norms may shift as long-term, multiple-turn agent use becomes common in contexts addressing social needs.
- Reframing the Concept of Conversation in Agent Interaction: Private and public settings shape what users consider appropriate to disclose or discuss with agents.
- Agent Conversation for Social Goals: Social robotics research suggests social capabilities can support elderly users’ disclosure of personal stories when interaction context and purpose are considered.
- Limitations and Future Work: The study did not involve participant interaction with conversational agents, and technology familiarity may have shaped reported views.
6 Conclusion
People viewed conversation as essential to human relationships but generally not important or desirable with current agents. The findings therefore point toward context-sensitive agent conversation that need not mirror human interaction.
- Participants described mutual understanding, common ground, trust, active listenership, and humour socially for humans but transactionally for agents.
- The findings suggest limits to how closely agent interactions can mirror conversations between people.
- Conversation may nevertheless be appropriate or essential in healthcare and wellbeing when context and demographics are considered.