Taxonomy of Human-AI Interaction Modes

• Taxonomy of human-AI interaction modes is a structured classification system that organizes collaborative workflows using formal primitives and design dimensions.
• It categorizes interactions by cognitive roles—replacement, augmentation, and extraheric—and control allocation, linking model operations to UI events.
• The taxonomy supports systematic design and evaluation across domains, from explainable AI to agent-based modeling, driving ethical, transparent system design.

A taxonomy of human–AI interaction modes specifies the structures, workflows, and intents that define collaborative activity between human agents and AI systems across a variety of domains, task horizons, and technical settings. Modern taxonomies leverage formal primitives, design dimensions, and empirical user studies to organize this space, supporting both rigorous comparative evaluation and principled system design. This article synthesizes major frameworks and dimensional taxonomies used in recent research, highlighting key modes, compositional hierarchies, and canonical design patterns.

1. Formal Primitives and Sequence Patterns

Contemporary taxonomies abstract human–AI interaction into formal primitives—atomic communication or operation types—combined into higher-order interaction patterns that map onto applications such as explainable AI, human-in-the-loop learning, and hybrid intelligence systems.

• Primitives: Provide (P) and request (R) actions, as formalized in message-passing models (e.g., $P\{X:\mathrm{Type}^+, Y:\mathrm{Type}^*\}$, $R\{X:\mathrm{Type}^+, Y:\mathrm{Type}^*\}$), where message type specifies the data object being exchanged (e.g., input sample, model output, feedback annotation).
• Patterns: Low-level primitives are sequenced into mid- and high-level interaction patterns (e.g., class selection, new sample annotation, prediction-based XAI, advice loop) that collectively define canonical workflows for supervised labeling, iterative model correction, reward shaping, and contestable AI negotiation loops.
• Interaction Modes: Patterns are grouped by functional purpose into high-level modes such as HITL/interactive ML, explainable AI (XAI), and hybrid intelligence, supporting fine-grained design reasoning across application domains (Tsiakas et al., 2024).

This compositional approach brings mathematical clarity, directly linking UI events and user communication to model- and operation-level semantics in deployed systems.

2. Cognitive and Functional Interaction Modes

Effective classification of human–AI interaction hinges on the functional role of AI relative to human cognition and action—replacement, augmentation, or stimulation of higher-order thinking.

• Replacement: AI automates full cognitive subtasks, minimizing intrinsic and extraneous cognitive load (e.g., real-time captioning, automatic translation, boilerplate code generation).
• Augmentation: AI enhances user cognition in defined subtasks without displacing human agency, optimizing efficiency for routine operators (e.g., autocomplete, script refactoring, query suggestion).
• Extraheric: AI intentionally increases germane cognitive load, soliciting creative, analytical, and critical thinking through scaffolding, questioning, viewpoint prompting, or debate, mapped to Analyze/Evaluate/Create levels of Bloom’s taxonomy (Yatani et al., 2024).

Each mode aligns with distinct load profiles and learning outcome targets, making this typology central to education, software development, and decision support design.

3. Autonomy and Oversight: Control Allocation Taxonomies

Interaction mode selection is profoundly influenced by the allocation of autonomy—who controls decisions, where human intervention is required, and at what risk threshold AI can act unilaterally. The six-mode taxonomy for technical services (Wulf et al., 18 Jul 2025) systematizes this spectrum:

Mode	Human oversight	AI autonomy	Key transition criterion
HAM	Full (human owns steps)	Minimal	Advisory only
HIC	Mandatory human approval	Low–Medium	Gating at approval
HITP	Embedded human subtask	Medium	Predefined workflow (no ad hoc escalation)
HITL	Exception escalation	Medium–High	Confidence-based trigger
HOTL	Discretionary supervision	High	Supervisor may intervene
HOOTL	None (post hoc only)	Very High	Complete automation for low-risk

Mode selection is formally regulated by task complexity $C$, operational risk $R$, system reliability $\rho$, and human operator’s workload $W$, e.g., $Mode = f(C, R, \rho, W)$. This model anchors human–AI system design against measurable project and organizational constraints.

4. Multidimensional, Domain-Specific Taxonomies

Recent taxonomies classify interaction modes using multi-axis design spaces, supporting precise characterization and gap analysis. Salient axes include:

• Interaction Intent, Degree, Focus, Feedback: The agency-interaction-adaptation model (Holter et al., 2024) expands interaction to a 4-tuple $(Intent, Degree, Focus, Feedback)$ per agent, where intent ranges from ReceiveGuidance to ProvideFeedback, degree from Orienting (multiple options) to Prescribing (single “best” path), focus (system/interface/data/task), and feedback modality (explicit/implicit). This formalism enables per-role interaction profiles in complex multi-agent systems.
• Control, Feedback, Engagement: Surveys of generative AI interaction extend to six axes: user purpose (e.g., refine, explore, automate), model feedback modalities, user control mechanisms (option selection, latent sliders, inpainting), engagement levels (passive–collaborative), domain, and evaluation strategy (Shi et al., 2023). Each system is a point in this high-dimensional interaction space, promoting systematic design and comparative evaluation.
• LLM-Driven Workflows: For LLM interfaces, a workflow divides into four canonical phases (Planning, Facilitating, Iterating, Testing) and four main interaction modes (Standard Prompting, UI-Augmented, Context-Based, Agent Facilitator), each characterized by the 5W1H schema (Who, What, When, Where, Why, How) (Gao et al., 2024).

In ABMS (agent-based modeling and simulation), a complete five-dimensional schema spans Why (goal: initialize/explore/refine/evaluate/analyze/immerse), When (pre/during/post-simulation), What (agents, environment, configuration), Who (scriptwriter/director/actor/prototype/observer), and How (interface, language, config, data, movement), offering a granular vocabulary for evolutionary simulation workflows and LLM integration (Lin et al., 25 Feb 2025).

5. Application-Specific Interaction Mode Taxonomies

Task or domain-centric taxonomies elaborate the interaction space for specific workflows:

• Text Summarization and Generation: Five canonical modes emerge: guiding model output (prompt, explicit parameters), selecting/rating outputs, post-editing, interactive editing (real-time, iterative), and writing with AI assistance. Modes further divide into open-loop (one-shot) and closed-loop (iterative/editing) workflows (Cheng et al., 2022).
• Software Engineering: Eleven mutually exclusive modes are identified, including auto-complete, command-driven actions, conversational assistance, contextual recommendations, selection-based enhancement, explicit UI actions, comment-guided prompts, event-based triggers, shortcut-activated commands, file-aware suggestions, and automated API responses. These modes are formally described by their trigger mechanism (automatic vs. explicit), locus of control, and output scope (inline, report, action) (Treude et al., 15 Jan 2025).

Mode Class	Example Trigger	AI Output	Canonical Example
Auto-complete	Typing	Inline suggestion	GitHub Copilot function
Event-based	Commit/PR	Diagnostic report	Auto DevOps scan
Conversational	Chat window	Explanations/code	ChatGPT in VS Code

These taxonomies provide foundational blueprints for tool designers and empirical research.

6. User Needs, Goals, and Conversational Appropriation

Taxonomies such as TUNA (Taxonomy of User Needs and Actions) ground interaction modes in empirically validated user action categories—Information Seeking, Processing/Synthesis, Procedural Guidance & Execution, Content Creation, Social Interaction, and Meta-Conversation—each further divided into mid-level strategies and atomic request types (e.g., explanation, summarization, error fix, persona directive). This tripartite structure enables cross-domain evaluation, interface policy harmonization, and detection of invisible or meta-conversational user labor (Shelby et al., 7 Oct 2025).

7. Implementation, Design, and Governance Strategies

A full implementation taxonomy articulates five categories—Human-Oriented Strategy Methods (value alignment, knowledge integration), Computation & Modelling (HCAI, explainability), Human Controllability (Meaningful Control, HITL design), Interaction Design (UI/UX, collaborative patterns), and Standards & Governance (audit, regulation)—situated within a hierarchical framework scaling from individual (Human-in-the-Loop) to organization, ecosystem, and society levels. This structure supports both system-level and policy-level harmonization, ensuring transparency, trust, and ethical alignment across application settings (Gao et al., 16 Jan 2026).

Category	Implementation Level	Examples
Human-Oriented Strategy	Individual/Org	Driver preference alignment, pilot co-train
Computation & Modelling	Individual/Org	Joint SA in AV, transparent cockpit
Human Controllability	Individual/Org	E-stop, approval gates
Interaction Design	Individual/Org	Heads-up HMI, cockpit explanation
Standards & Governance	Ecosystem/Society	SAE, FAA, FDA regulation

This integrated approach bridges the gap between technical design and normative social/legal constraints.

---

References:

• (Tsiakas et al., 2024): Unpacking Human-AI interactions: From interaction primitives to a design space
• (Yatani et al., 2024): AI as Extraherics: Fostering Higher-order Thinking Skills in Human-AI Interaction
• (Wulf et al., 18 Jul 2025): Architecting Human-AI Cocreation for Technical Services -- Interaction Modes and Contingency Factors
• (Cheng et al., 2022): Mapping the Design Space of Human-AI Interaction in Text Summarization
• (Holter et al., 2024): Deconstructing Human-AI Collaboration: Agency, Interaction, and Adaptation
• (Shi et al., 2023): An HCI-Centric Survey and Taxonomy of Human-Generative-AI Interactions
• (Lin et al., 25 Feb 2025): Carbon and Silicon, Coexist or Compete? A Survey on Human-AI Interactions in Agent-based Modeling and Simulation
• (Gao et al., 16 Jan 2026): Toward Human-Centered Human-AI Interaction: Advances in Theoretical Frameworks and Practice
• (Gao et al., 2024): A Taxonomy for Human-LLM Interaction Modes: An Initial Exploration
• (Treude et al., 15 Jan 2025): How Developers Interact with AI: A Taxonomy of Human-AI Collaboration in Software Engineering
• (Shelby et al., 7 Oct 2025): Taxonomy of User Needs and Actions