Hierarchical Co-Creative System

• Hierarchical co-creative system is a structured framework that balances human creativity and AI autonomy through clearly defined levels of interaction.
• It organizes workflows, responsibilities, and information flows using formal mappings and synchronization protocols to maintain creative consistency.
• Empirical evidence from domains like mathematics and multimedia demonstrates its potential to accelerate exploration and overcome creative barriers.

A hierarchical co-creative system is a structured framework for collaborative interaction between humans and AI systems, where both parties contribute to the creative process, but with clearly differentiated levels of autonomy, agency, and interaction richness. This architecture organizes workflows, responsibilities, and information flows to enable scalable, context-aware, and fluid co-creation across domains such as mathematics, design, creative writing, and multimedia production (Haase et al., 19 Nov 2024, Crowley et al., 2022, Cao et al., 6 Mar 2025, Lin et al., 2023).

1. Hierarchical Models of Human–AI Co-Creativity

The dominant paradigm for hierarchical co-creative systems involves a stratified organization of interaction levels, with each level characterized by incremental degrees of AI autonomy, creative agency, and communication complexity. In the four-level model formalized by Haase & Pokutta (2024), the system hierarchy comprises:

• Level 1: Digital Pen (no autonomous AI generation; tool serves as digitized CSS, all creativity is human-driven).
• Level 2: AI Task Specialist (AI fully automates narrowly defined subtasks under explicit parameters, returning batch-mode results subject to human review).
• Level 3: AI Assistant (general-purpose generative models provide real-time, guided creative suggestions via interactive chat/prompt-based sessions).
• Level 4: AI Co-Creator (human and AI operate as near-equal creative partners, jointly seeding, evolving, and refining novel solutions through dynamic, bidirectional dialogue) (Haase et al., 19 Nov 2024).

Each successive level exhibits an increased degree of autonomy $A(L_i)$, creative agency $C(L_i)$, and richness of interaction $R(L_i)$, e.g., $A(L_1)<A(L_2)<A(L_3)<A(L_4)$.

Extended ontologies elaborate these roles into “computer-as-subcontractor,” “computer-as-critic,” and “computer-as-teammate” (with further peer/apprentice/master distinctions), focusing on the division of creative labor, information flow, and initiative (Lin et al., 2023).

2. Formal Structures, Protocols, and Substrates

The hierarchical organization is instantiated via compositional structures and explicit mappings among them, as exemplified in the VideOrigami environment (Cao et al., 6 Mar 2025). These structures facilitate multiple granularities of content creation, planning, and refinement:

Structure	Function	Atomic Units (U)
Canvas (C)	Spatial repository (assets, notes, prompts)	Nodes
Narrative (N)	Linear, block-based editing	Sections, Paragraphs
Grid Planner (G)	Tabular scene/task planning	Cells (Scene x Track columns)
Timeline (T)	Temporal sequencing (multi-track)	Snippets (audio, visual, text)

Mappings $\varphi$ (e.g., $\varphi_{NG}: U_N \rightarrow G_{rows}$) ensure that edits and ideas propagate consistently across structures. Within and across these substrates, AI is infused to assist with content generation, summarization, visual previews, and audio transcription, but remains constrained by explicit organizational rules and synchronization logic.

Communication protocols reflect the level of initiative and information required. L1 relies on direct manipulation in GUI paradigms; L2 exchanges parameterized specs for batch execution; L3 leverages iterative chat or prompt-response; L4 uses meta-controller-driven, mixed-initiative dialogue with adaptive calibration to user expertise (Haase et al., 19 Nov 2024, Cao et al., 6 Mar 2025).

3. Allocation of Agency, Authority, and Initiative

Role allocation and initiative functions are rigorously formalized in hierarchical co-creative systems. Agents $A_i = (\Sigma_i, B_i, M_i)$ (sensor channels, actions, internal model) interact according to a fixed or adaptive authority function $\alpha: Tasks \times Agents \rightarrow [0,1]$, partitioning decision rights and control (Crowley et al., 2022).

• Computer-as-Subcontractor: AI executes human-assigned subtasks autonomously.
• Computer-as-Critic: AI provides feedback on human drafts but lacks generative initiative.
• Computer-as-Teammate: Mixed-initiative exchange, with turn-taking and initiative triggers (peer, apprentice, or master mode).

Information flow models span one-shot (subcontractor), iterative (critic), and continuous, mixed-initiative interactions (teammate), supporting granular responsibility division and artifact state evolution (Lin et al., 2023). In high-level co-creative settings, AI must dynamically adapt its initiative, posing clarifying queries, proposing alternate strategies, or shifting between “explore” versus “exploit” behavior via meta-calibration modules—especially in L4 (Haase et al., 19 Nov 2024).

4. Algorithms, System Architectures, and Synchronization

The architectural pipeline is tailored to the level of creativity and required interaction. L1 typically involves non-ML, client–server I/O stacks (collaboration tools). L2 introduces specialized optimization engines (e.g., Frank–Wolfe solvers), with computation decoupled from interface interaction and escalation strictly user-triggered. L3 generalizes to pipelines with pretrained LLMs or diffusion backbones, interfaced via chat/prompt encoders and constrained by user-driven guardrails. L4 adds a meta-controller atop L3, monitoring dialogue, satisfaction, and adapting escalation logic (Haase et al., 19 Nov 2024).

In compositional environments, synchronization and aggregation mechanisms coordinate edits and ideas bidirectionally. Synchronized highlighting and editing, guided by the $\varphi$ mappings between structure units, enable users to persist orientation, propagate context, and prevent destructive overwrites. UI mechanisms such as global state dispatch/subscription, highlighting, and undo stacks maintain synchrony across multiple concurrent substrates (Cao et al., 6 Mar 2025).

5. Empirical Evidence, User Evaluation, and Design Principles

Hierarchical co-creative systems have demonstrated empirical efficacy across mathematical and creative domains. Key evidence includes:

• Mathematics: L2 systems found new Bell inequalities in previously unstudied scenarios (reported ten-fold increase in exploration speed); L3 systems helped derive tighter upper bounds on Ramsey multiplicities for K₄ using AI-generated constructions; L4 systems proposed new six-color tilings that broke long-standing combinatorial barriers—outputs confirmed by human experts and verified formally (Haase et al., 19 Nov 2024).
• Video Co-creation: The VideOrigami study showed that explicit compositional substrates prevented “prompt fatigue,” promoted project-wide awareness, and enabled both rapid iteration and precise user control. Novice and expert evaluation favored grid and timeline planning, with “automation utility” rated highest for synchronization and generation, though experts noted tensions regarding creative control cession. Emergent workflows often centered on switching among several structures, surfacing novel trade-offs between efficiency and creative evaluation cost (Cao et al., 6 Mar 2025).

General design principles distilled from user studies and architectural analysis include:

1. Expose key domain structures as interactive scaffolds—spatial, narrative, congruence, temporal—for flexible choice and navigation.
2. Define explicit cross-structure correspondences and aggregation mechanisms to propagate ideas and changes while controlling information directionality.
3. Integrate context-aware AI that respects both structural constraints and user intent—within substrates and across linked representations.
4. Balance automation with human agency by employing “suggest-and-accept” patterns, transparent prompts, confidence indicators, and selective propagation regimes.
5. Support freeform exploration, versioning, and recovery through canvas or sketch layers and persistent, lightweight revision histories (Cao et al., 6 Mar 2025, Haase et al., 19 Nov 2024).

6. Open Challenges and Prospects

Persistent open research challenges include sample-efficient learning of new creative affordances, real-time personalization of explanations, and long-term trust/alignment maintenance between human and machine partners. The formalization of communication, explanation, and situation model alignment remains a focal point for robust co-creative system design (Crowley et al., 2022). There is particular emphasis on ethical transparency—explicitly communicating the AI’s initiative and role (subcontractor, critic, or teammate)—to foster user trust and manage expectations (Lin et al., 2023).

The hierarchical co-creative paradigm is extensible to emergent domains such as personalized robotic services, education, accessible multimedia creation, and collaborative knowledge work. Its ability to scale creative contribution without diminishing human intent or expertise is grounded in a rigorously structured allocation of agency, initiative, and information flow—anchored in both empirical results and extensible, compositional architectures (Haase et al., 19 Nov 2024, Cao et al., 6 Mar 2025, Crowley et al., 2022, Lin et al., 2023).