Taxonomy of Human-AI Interaction

• Taxonomy of Human-AI Interaction Patterns is a comprehensive framework that categorizes the protocols, design choices, and communication mechanisms between humans and AI systems.
• It employs multidimensional structures such as the 5W1H formalism and role-based paradigms to classify interaction elements like agency distribution, control modalities, and feedback mechanisms.
• The framework identifies current limitations and gaps, guiding research towards enhanced collaboration, adaptive learning, and transparent evaluation in human-AI systems.

Taxonomy of Human-AI Interaction Patterns

Human–AI interaction patterns codify the mechanisms, protocols, and design choices by which humans and AI systems collaborate, communicate, or allocate decision authority in complex tasks. Multiple taxonomies, developed through systematic reviews and empirical user studies, span granular message-passing models to overarching paradigms of co-agency and cognitive trade-off, providing both formal notations and multidimensional frameworks. This article systematically synthesizes these taxonomies, focusing on rigorous definitions, classification schemes, underlying theoretical foundations, and design implications relevant to technical audiences.

1. Multidimensional Structure of Human–AI Interaction Taxonomies

Core taxonomies in the literature decompose human–AI interaction into multiple orthogonal or quasi-independent dimensions. A dominant organizing structure is the 5W1H formalism, as exemplified for agent-based modeling and simulation (Lin et al., 25 Feb 2025):

• Why: The user's fundamental intent (e.g., initialization, exploration, refinement, evaluation, analysis, immersion)
• When: Temporal locus of interaction (pre-simulation, during simulation, post-simulation)
• What: Target components for control (agents, environment, simulation configuration)
• Who: User role or viewpoint (scriptwriter, director, actor, prototype, observer)
• How: Interaction modalities (GUI, natural language, configuration files, data integration, physical movement)

Different domains instantiate this multidimensional approach using tailored dimensions. For example, Holter & El-Assady (Holter et al., 2024) define:

• Agency: Distribution (human, AI, mixed); allocation (pre-determined, negotiated)
• Interaction: Intent (guidance, info request, exploration, feedback); degree/focus of guidance; feedback type (explicit, implicit)
• Adaptation: Which agents adapt; method (task vs. communication improvement); information learned (domain, data, task, goals, preferences)

This multidimensionality enables comparative analysis, facilitates the identification of underexplored design spaces, and supports modular system engineering.

2. Representative Taxonomic Schemes and Paradigms

2.1 Workflow-Centric and Pattern-Based Taxonomies

Several taxonomies operationalize human–AI interaction patterns via concrete workflow patterns, most notably for text summarization and generation (Cheng et al., 2022):

Pattern	Human Action	Control Placement	In-Loop Iteration	Initiator	Application Domains
Guiding Model Output	configure	pre-generation	no	human	Summarization, NLG
Selecting/Rating Output	select/rate	post-generation	no	AI	Summarization, MT
Post-Editing	edit	post-generation	no	AI	Summarization, creative
Interactive Editing	edit+AI regen	both	yes	both	MT, Summarization
Writing w/ Assistance	author/accept	ongoing	yes	human	Writing, coding, creative

Patterns are compared along efficiency, control, and trust axes, and design recommendations emphasize adjustable assistive granularity, transparent provenance, and always-available post-editing.

2.2 Human Feedback and Alignment Reference Models

Human–AI alignment research distinguishes the interaction pipeline into five operator stages, forming a taxonomy of UI contributions (Shi, 12 Feb 2026):

1. Data Sampling (S): Selection from all possible outputs (random, active, clustered)
2. Visual Transformation (T): Mapping outputs to visualizations (text, t-SNE, graphs)
3. Interactive Display (D): Configuring the display's interactive elements
4. Analysis Interaction (α): User-driven comparison, exploration, reordering
5. Feedback (F): Signal type (binary, ranking, rating, free-form)

Each user interface is a concrete instantiation in this five-dimensional design space, and changing any dimension yields a new interaction pattern (e.g., groupwise vs. pairwise comparison). Formal notation:

$$S: O \to P(O); \quad T: P(O) \to V; \quad D: V \to U; \quad \alpha: U \to A; \quad F: A \to P$$

3. Agency, Collaboration, and Role Allocation

Taxonomies frequently analyze how agency is allocated and how responsibility for goal definition, control, and learning is distributed.

• Agency Distribution: Holter & El-Assady (Holter et al., 2024) classify systems as human-centric, AI-centric, or mixed.
• Levels of Agency Allocation: Pre-determined (fixed authority) vs. negotiated (dynamic, context-adaptive).
• Collaboration Levels in RL: Augmentative (one augments, the other leads), Integrative (shared task, complementary skills), Debative (disagreement and negotiation) (Li, 2024).
• Responsibility in Creative Domains: Shi & Choo (Shi et al., 25 Feb 2026) classify design ideation into Human-Only, Human-Lead, AI-Lead, Co-Evolution, based on relative creative responsibility.

Allocation of role and task strongly affects interaction design, trust calibration, and system flexibility.

4. Interaction Modalities, Control, and Feedback Mechanisms

A recurring axis is the modality and explicitness of user–AI exchanges.

• Interaction Primitives: “Provide” and “Request” of specific information types form the basis of all exchange (Tsiakas et al., 2024). Patterns are created by compositional sequences of these primitives, such as sample annotation, advice, explanation request, modification, and recommendation.
• Control Modalities: Ranges from explicit commands and parameter adjustments (e.g., prompt engineering, sliders), to selection (choosing among proposals), to direct manipulation (drag-and-drop, brush inpainting), as detailed in (Luera et al., 2024, Shi et al., 2023).
• Feedback Types: Explicit (direct input or labeling), implicit (interaction traces), or both (Holter et al., 2024). Feedback encompasses confidence signals, ratings, provenance links, and explanations.

Interaction complexity and user autonomy are linked to task risk, user expertise, and system explainability.

5. Evaluation Axes and Trade-Offs

Multiple taxonomies formalize evaluation criteria and trade-offs:

• Efficiency, Control, Trust: Cheng et al. position interaction patterns in a vector space for these qualities (Cheng et al., 2022).
• Cognitive Load Theory: Extraheric AI taxonomy (Yatani et al., 2024) classifies system types by how they redistribute intrinsic, extraneous, and germane cognitive load.
• Contingency-Based Mode Selection: Wulf et al. (Wulf et al., 18 Jul 2025) connect interaction mode (HAM, HIC, HITP, HITL, HOTL, HOOTL) to measured system reliability, task complexity, and operational risk, using explicit scoring functions for autonomy.
• Functional Capabilities: Dafoe et al. (Li, 2024) decompose collaborative capacity into understanding, communication, commitments, and institutional frameworks.

Patterns are selected or constructed based on context, with guidelines aligned to these dimensions.

6. Limitations, Open Questions, and Research Directions

Consensus across taxonomies identifies several limitations in current practice:

• Interaction patterns are dominated by simplistic or unidirectional paradigms, often neglecting bidirectional, negotiated, or longitudinal adaptation (Gomez et al., 2023, Holter et al., 2024).
• Agency allocation remains under-theorized, especially regarding dynamic negotiation and co-adaptive learning.
• Few studies operationalize the impact of interaction modality, control synchronization, and feedback richness on metrics such as trust and cognitive load (Cheng et al., 2022, Yatani et al., 2024).
• Ethics, transparency, and responsible AI governance are insufficiently mapped to interaction design choices.
• Multimodal and multi-agent collaboration, the evolution of user roles, and the measurement of germane cognitive engagement require further empirical and formal study.

Research directions include the formalization of rigorously measurable interaction affordances, the co-evolutionary modeling of user and AI learning, and the development of adaptive, context-sensitive interaction frameworks.

---

Taxonomies of human–AI interaction patterns supply a systematic, multi-axial vocabulary and set of formal mappings enabling the analysis, design, and evaluation of collaborative, assistive, and co-creative systems across domains and modalities. Contemporary work advances the field by bridging low-level interaction primitives to high-level paradigmatic trade-offs, offering practical templates and formal metrics to optimize human–AI efficacy, adaptation, and trustworthiness.