Generative AI Agents for Controllable and Protected Content Creation
Haris Khan, Sadia Asif · 2026
Read Generative AI Agents for Controllable and Protected Content Creation by Haris Khan, Sadia Asif (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-12348.txt.
Introduction
Haris Khan and Sadia Asif build generative AI agents for content that users can control and that stays protected. Generative tools have changed creative work. Yet users lack fine control over output, and output lacks proof of origin once made.
Their multi-agent frame splits the work among agent roles with set tasks. It adds watermarking in the generation step itself, not as an afterthought. Past multi-agent tools chase quality alone. A built-in mark plus control could aid trust and rights in AI media. Marks made at birth of a work are hard to strip.
Catalog record
| CREATOR | Haris Khan, Sadia Asif |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 29 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-12348
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Generative AI Agents for Controllable and Protected Content Creation Haris Khan, Sadia Asif Abstract The proliferation of generative AI has transformed creative workflows, yet current systems face critical challenges in controllability and content protection. We propose a novel multi-agent framework that addresses both limitations through specialized agent roles and integrated watermarking mechanisms. Unlike existing multi-agent systems focused solely on generation quality, our approach uniquely combines controllable content synthesis with provenance protection during the generation process itself. The framework orchestrates Director/Planner, Generator, Reviewer, Integration, and Protection agents with human-in-the-loop feedback to ensure alignment with user intent while embedding imperceptible digital watermarks. We formalize the pipeline as a joint optimization objective unifying controllability, semantic alignment, and protection robustness. This work contributes to responsible generative AI by positioning multi-agent architectures as a solution for trustworthy creative workflows with built-in ownership tracking and content traceability. \workshoptitle The First Workshop on Generative and Protective AI for Content Creation Keywords: Generative AI, Multi-agent systems, Content protection, Controllability, Watermarking, Creative AI pipelines 1 Introduction The rapid advancement of generative artificial intelligence has revolutionized content creation across text, image, and video domains. Systems like GPT-4 gpt4 , DALL-E dalle2 , and Stable Diffusion stablediffusion enable unprecedented creative possibilities, transforming how content is produced in professional and commercial contexts. However, two fundamental challenges limit the adoption of these technologies in real-world applications. First, current generative systems often operate as “black boxes,” providing limited controllability over output characteristics beyond initial prompts promptdesign . Users struggle to achieve precise alignment between their creative intent and generated content, particularly for complex, multi-faceted creations requiring specific compositional elements, stylistic choices, or semantic constraints. This controllability gap becomes critical when precision and iterative refinement are essential. Second, generated content lacks inherent protection mechanisms, creating risks for copyright infringement, unauthorized usage, and content provenance tracking diffusionart . As generative AI outputs become increasingly sophisticated and difficult to distinguish from human-created content, the need for embedded protection and traceability mechanisms grows urgent. Current watermarking approaches typically apply protection as a post-processing step, which can degrade quality and reduce robustness stablesignature . While multi-agent frameworks have emerged for complex generative tasks mulan ; animaker , existing approaches focus primarily on improving generation quality rather than addressing controllability and protection jointly. Recent systems like MetaGPT metagpt and ChatDev chatdev demonstrate effective role-based collaboration for software development, while TradingAgents tradingagents and NovelSeek novelseek apply multi-agent architectures to financial modeling and literature search respectively. However, none integrate protection mechanisms directly into the generation pipeline or provide the fine-grained human-in-the-loop control necessary for creative applications. To address these gaps, we introduce a multi-agent generative framework that decomposes creation into controllable stages and embeds protection directly into the pipeline. Our key innovation lies in the joint optimization of controllability and protection objectives through specialized agent collaboration. The Planner and Reviewer agents provide iterative alignment between user intent and system output, while the Integration and Protection agents ensure coherent composition and imperceptible watermarking for provenance tracking—all within a unified generation loop rather than as separate post-processing steps. This paper makes three key contributions: (1) We propose a novel multi-agent generative pipeline that integrates protective AI mechanisms directly into controllable content creation; (2) We formalize the framework through a joint optimization objective that unifies planning, generation, review, integration, and protection; (3) We outline a comprehensive experimental evaluation plan with quantitative metrics and ablation studies to validate the framework’s effectiveness. 2 Related Work 2.1 Multi-Agent Generative Systems Recent work has demonstrated the effectiveness of multi-agent approaches for complex generative tasks. MuLan mulan pioneered the integration of large language model planning with diffusion models for compositional image generation, showing how agent-based decomposition can improve controllability through structured task breakdown. AniMaker animaker extended this concept to video generation, employing specialized agents including Directors for scene planning, Reviewers for quality assessment, and Post-Production agents for final refinement. Beyond creative generation, multi-agent frameworks have proven successful in software development and specialized domains. MetaGPT metagpt introduced a collaborative framework where agents assume roles such as product managers, architects, and engineers to produce structured software artifacts. Similarly, ChatDev chatdev demonstrated how role-playing agents can iteratively refine code through simulated organizational structures. TradingAgents tradingagents applied multi-agent collaboration to financial modeling, while NovelSeek novelseek employed agent cooperation for literature search and knowledge discovery. While these systems demonstrate that role-based decomposition improves task performance, existing approaches focus on either generation quality or task-specific optimization. None integrate content protection mechanisms into the generation loop itself, nor do they provide the fine-grained human-in-the-loop control necessary for creative applications where iterative refinement and ownership tracking are paramount. 2.2 Generative AI and Information Retrieval The intersection of generative AI and information retrieval (IR) has produced significant innovations, particularly in retrieval-augmented generation (RAG) systems rag ; fid . These approaches combine retrieved knowledge with generative capabilities to produce contextually relevant content. Recent work has explored multimodal RAG systems that generate images, audio, and video content informed by retrieved information multimodalrag ; videorag . Modern IR workflows increasingly involve not just retrieving existing content, but generating new content tailored to user needs rethinkingsearch ; beir . This evolution demands systems that maintain the precision and trustworthiness traditionally associated with IR while embracing generative flexibility. Our framework contributes to this intersection by providing controllable generation mechanisms that align with IR community values of relevance, precision, and verifiable provenance. 2.3 Protective AI and Content Provenance The growing sophistication of generative AI has intensified focus on content protection mechanisms. Recent advances in watermarking for diffusion models show promise for embedding imperceptible markers in generated images stablesignature ; watermarkrecipe . Chen et al. robustwatermark demonstrated robust watermarking techniques that survive common image transformations while maintaining generation quality through adversarial training. Beyond watermarking, researchers have explored fingerprinting techniques for model identification and provenance tracking artificialfingerprint ; pii . These approaches enable post-hoc detection of AI-generated content and attribution to specific generation systems. Despite these advances, existing protective mechanisms are typically applied as post-processing steps rather than being integrated into the generation pipeline. This separation can lead to quality degradation and reduced protection robustness. Our framework uniquely addresses this limitation by embedding the Protection agent directly within the multi-agent loop, enabling protection-aware generation that maintains quality while ensuring robust watermark embedding. 3 Proposed Multi-Agent Framework 3.1 Architecture Overview Our proposed framework orchestrates five specialized agents in a sequential pipeline designed to maximize both controllability and protection.Our contribution differs from prior multi-agent systems by (i) integrating a Protection agent within the generation loop—changing both the optimization objective and data flow; (ii) jointly optimizing controllability and watermark robustness in Eq. (10), rather than treating protection as post-hoc; and (iii) exposing protection parameters to the human-in-the-loop controller for traceable creative decisions. Figure 1 illustrates the complete workflow, emphasizing the iterative nature and human intervention points that enable fine-grained control over the generation process. The framework operates on the principle of task decomposition, where complex generative requests are broken down into manageable subtasks handled by domain-specific agents. This modular approach enables targeted optimization of each generation stage while maintaining overall coherence through the Integration agent and ensuring protection through watermark embedding during synthesis rather than post-processing. Figure 1: Proposed multi-agent framework. The pipeline consists of specialized agents for planning, generation,