Understanding Human-Human vs. Human-AI Interaction

With artificial intelligence (AI) chatbots being rapidly integrated into various sectors of life, such as customer service, healthcare, and education, there’s a growing need to understand their potential and limitations in replicating human interaction. This has been a hot topic among practitioners in conversational AI and conversation analysis, two interconnected disciplines that enrich our understanding of human vs. human-AI interaction.

Conversational AI

IBM defines conversational AI as “technologies, such as chatbots or virtual agents, that users can talk to. They use large volumes of data, machine learning, and natural language processing to help imitate human interactions, recognizing speech and text inputs and translating their meanings across various languages.” Notable examples of conversational AI are Apple’s Siri and Amazon’s Alexa, both of which provide services by “conversing” with users. The key components of conversational AI are machine learning (ML), natural language processing (NLP), natural language understanding (NLU), text-to-speech (TTS), speech recognition, and reinforcement learning (RL).

• Machine learning (ML) – a sub-field of AI that uses adaptive algorithms to accurately recognize patterns and make predictions as the amount of input grows.
• Natural language processing (NLP) – an ML–driven approach for analyzing language in conversational AI.
• Natural language understanding (NLU) – a subset of NLP focused on comprehending the meaning behind user inputs, including intent and sentiment.
• Text-to-speech (TTS) – converts text to audible speech for human listeners.
• Speech recognition – converts spoken language into text so the system can process voice inputs.
• Reinforcement learning (RL) – ML algorithms for refining responses to improve accuracy through feedback (rewards or penalties).

Conversation Analysis

Conversation analysis (CA) was developed by sociologists Harvey Sacks, Gail Jefferson, and Emmanuel Schegloff in the 1970s as a systematic tool to uncover sequence organization that constitutes the interactional order. CA has been applied to study a variety of interactions, including doctor-patient interaction, instructional interaction (e.g., driving lessons, classroom discourse, dance lessons), media interaction, and so on. While the initial focus has been on talk-in-interaction, CA has been expanded to the studies of multimodal interaction (how silent actions, objects, and bodies are used in interaction).

As conversational AI continues to evolve, its ability to mimic human interaction has become increasingly sophisticated. However, the question remains: To what extent can AI truly replicate the nuances of human conversation? This is where insights from CA become invaluable. CA can be used as a tool to uncover the nuances of human-human interaction as well as to improve the state-of-the-art human-computer or human-robot interaction.

Comparing Human-Human and Human-AI Interaction

The differences between human-human and human-AI interaction have been studied by applied linguists Stokoe et al. (2024) and Voss and Waring (2024). Here’s a brief overview of these differences based on the researchers’ interaction with ChatGPT-3.5.

Turn-taking – the mechanism by which speakers alternate speaking turns in a conversation. Turns by humans consist of turn constructional units (TCUs), each ending with possible transition relevance places (TRPs) that allow the other speaker to jump in. In contrast, AI produces massive multi-TCUs without TPRs, overlaps, gaps, or pauses—like a professor going on and on in a lecture!

Action formation and recipient design – how speakers craft utterances to accomplish specific social actions (action formation), and how they tailor talk to their audience (recipient design). Human interaction is not limited to service interaction, and humans use a lot of interactional resources to get their point across depending on the social activity. By contrast, AI thus far has been preset to provide service (e.g., “How may I assist you?”). Its repertoire is service-oriented.

Activities – different social actions people perform when communicating. Humans engage in tons of activities on a daily basis, including arguing, playing, questioning, flirting, resisting, and so on. On the other hand, given AI’s orientation to provide service, its activities are mainly limited to informing, searching, answering, and advising.

Structural organization – the overall shape and flow of conversation. Human-human interactions tend to be more natural and egalitarian, whereas AI tends to be lengthy and “lecture-like.”

Indexicality – using context-bound expressions and references that rely on shared context. Human interactions are highly indexical. They use cohesive devices and discourse markers to reference prior talk. In contrast, AI lacks orientation to indexicality (e.g., likes to repeat a whole chunk of text).

Takeaways

As conversational AI continues to play a critical role in various sectors, CA can serve as a useful tool for improving the naturalness of human-AI interaction. This is especially important as we begin to witness the emergence of robotics (physical AI) in everyday life. Rather than humans beginning to sound like robots, robotics interaction should be developed based on the naturalness of human-human interaction to better assist all walks of life. Ultimately, the success of conversational AI will depend on its ability to navigate the complexities of human dialogue.

Note: This post is written by a human with some AI assistance 🙂. The author is not sponsored by any organizations or persons mentioned.