Human-AI Interaction Card

• Human-AI Interaction Card is a framework that structures dialogue design to manage user cognition and trust in conversational interfaces.
• It applies empirical research and cognitive science to optimize phrasing, speech modulation, and expectation management, reducing cognitive load.
• The design guidelines include key sections like header, prompt guidance, and transparency indicators to support effective Theory of Mind development.

Human–AI Interaction Card

Human–AI Interaction Cards serve as structured, actionable scaffolds for the design, evaluation, and deployment of conversational and natural language interfaces. Drawing from empirical survey research, cognitive science, and interface design methodologies, Card frameworks provide empirically grounded templates to guide end-user interactions, manage user cognition and trust, and reduce excess cognitive effort when communicating with AI systems in text and voice modalities (Adkins, 2024).

1. Definition and Cognitive Science Foundations

Human–AI Interaction (HAI) is a subfield of Human-Computer Interaction (HCI) that investigates how people communicate intent, receive feedback, and build trust with intelligent agents. Natural Language Interfaces—encompassing both textual chat and voice assistants—aim to let users interact as they would with another human, without recourse to rigid menu-driven or command-based protocols. However, empirical results demonstrate that current interfaces impose substantial cognitive overhead due to the restricted communication channels of state-of-the-art AI models (Adkins, 2024).

At the cognitive level, effective HAI depends critically on the user’s Theory of Mind (ToM)—the human ability to form predictive models of a dialogue partner’s knowledge, capabilities, and likely responses. In human-to-human conversation, mutual ToM is established and stabilized via explicit and multimodal signals (intonation, gesture, facial expression), enabling fluid, low-effort exchanges. By contrast, HAI interactions are constrained to a one-sided ToM: users strive to infer AI capabilities in the absence of nonverbal feedback, leading to persistent uncertainty, non-fluid turn-taking, and requirement for systematic, often unnatural, mental modeling of “how to speak so the AI will understand” (Adkins, 2024).

2. Interaction Patterns and Cognitive Implications

Empirical studies delineate three prototypical interaction patterns in text or voice-based HAI systems:

• Text Chat Adaptation: Users systematically alter phrasing—shortening/lengthening utterances and excising natural filler words—to optimize parsing by the AI. This adaptation produces measurable planning overhead and reduces conversational fluidity.
• Speech Modulation: In vocal interfaces, users slow speech, enunciate carefully, choose simpler constructions, and modulate their volume. Resultant exchanges are stilted, and attention is diverted from the task domain to “talking like a machine.”
• Contrast With Peer Dialogue: Users rarely talk to AI systems as they would to friends; they report greater attention to formality and structure, and find HAI dialogue persistently unnatural, prompting higher working memory demands and frequent utterance replanning.

These patterns are associated with increased cognitive load, as evidenced by qualitative and quantitative measures. For example, users report the need to “think systematically” and indicate lower conversational naturalness and fluency compared to human–human interaction (Adkins, 2024).

3. Quantitative Measures and Experimental Design

The referenced study utilized N=101 adult participants employing a 25-item Likert-scale survey, supplemented with frequency, yes/no, multi-select, and free response items. For each item $j$, Likert-weighted responses were scored as $M_j = (1/N) \cdot \sum_{i=1}^N w_{i,j}$, mapping “strongly disagree” to –2 through “strongly agree” to +2. Specific metrics include:

• Repetition/Rephrase Frequency: Self-reported necessity to repeat or rephrase instructions, yielding an average rephrasing score $R$ (Never=1 … Always=5).
• Confidence Scores: For voice systems, confidence in whether the AI “heard” or “understood” the user, mapped on a –2…+2 scale.
• Task Efficiency: Inferred from lower repetition/rephrase counts and self-reported conversational cycle speeds.
• Cognitive Load Proxy: Indirectly measured via qualitative responses to items such as “I have to think more systematically” (Adkins, 2024).

Table: Mean Ratings for Representative Theory of Mind Items

Survey Item	Mean Score
“I change how I phrase things to AIs…”	+1.02
“Interacting with an AI feels natural”	–0.57
“I phrase things with an AI as with a friend”	–1.00

4. Main Findings on Theory of Mind in HAI

Users do actively attempt to develop a theory of mind for AI agents—demonstrated by consistent adaptation of phrasing and delivery—but typically fail to achieve interactional comfort or fluency. The absence of visible model reasoning and nonverbal feedback cues enforces a “one-way stranger” configuration, fundamentally distinct from mutual and embodied human dialogue.

Crucially, this leads to a ToM that remains persistently immature: users continue to invest cognitive effort in modeling the AI but do not realize the expected benefits of improved ease or efficiency with experience. Increased cognitive load persists over time rather than diminishing, in contrast with the learning effects observed in expert–novice human communication (Adkins, 2024).

5. Card Structure and Design Guidelines

A Human–AI Interaction Card should contain the following canonical sections to support user understanding, expectation management, and cognitive load reduction:

1. Header
   • Title: e.g., “How to Talk with [Agent Name]”
   • Description: Clear statement of reference use (“Your quick reference for phrasing, expectations, and tips…”)
2. Definition Block
   • Strengths: Concise statement of what the agent excels at (e.g., “optimized for factual questions and step-by-step instructions”)
   • Limitations: Explicit statement of agent boundaries (“cannot recognize sarcasm or personal context”)
3. Prompt and Phrasing Guidance
   • Text Templates: e.g., “Summarize X in three bullet points.”
   • Voice Tips: e.g., “Speak clearly, pause after each clause.”
4. Transparency and Explainability
   • Confidence Indicators: e.g., “I’m 80% certain I understood you.”
   • Explanation Access: “Why/how” buttons or hover text.
   • Version Stamp: e.g., “Model updated May 2024.”
5. Expectation Management
   • Boundary Reminders: e.g., “I treat every question literally.”
   • Timing: e.g., “Most replies within 2 seconds.”
   • Correction Workflow: e.g., “To refine my answer, say ‘That’s not what I meant. I wanted…’”
6. Cognitive Load Management
   • Refinement Pointers: e.g., “You can say ‘Add more detail.’”
   • Chunking: e.g., “Ask one question at a time for best results.”
   • Memory Limit: e.g., “I only remember the last two prompts.”
7. Footer: Feedback and Metrics
   • Mini-survey: e.g., “Was this answer helpful?” Yes/No.
   • Error Reporting: Explicit feedback contact.
   • Cognitive Load Tip: e.g., “If you find yourself rephrasing a lot, try rephrasing to one simple question.”

This template is intended to be visually scannable, concise, and integrated as an active design artifact supporting end-user ToM development for AI agents (Adkins, 2024).

6. Broader Implications and Future Directions

The Human–AI Interaction Card framework addresses persistent limitations in current conversational and natural language interfaces by providing explicit scaffolding for expectation management, explanation, and user-cognitive adaptation. Given the empirical finding that users’ mental models do not stabilize and cognitive load does not attenuate naturally over extended use, the inclusion of structured, transparent, and explicit interface guidance is essential.

Long-term implications for interface and workflow design include the need for continual tuning of prompts and transparency mechanisms, richer explainability tailored to evolving user models, and adaptive strategies for load reduction. Design teams are advised to monitor rephrasing rates and user feedback, iteratively updating card elements to incrementally reduce systematic user workarounds, with the ultimate goal of approximating the comfort and effortlessness of mature human–human communication in the context of AI-mediated dialogue (Adkins, 2024).