FoldMark: Watermarking Frameworks for Embedding, Detecting, and Authenticating Edits in AI-Generated Content

Background

Large language models (LLMs) have become foundational technologies across diverse domains, including education, software development, and customer service, due to their ability to generate human-like text. As these models are increasingly used to produce content at scale, concerns have arisen regarding the authenticity and integrity of AI-generated text. Watermarking techniques have emerged as a solution to distinguish between machine-generated and human-authored content, supporting efforts in copyright enforcement, misinformation mitigation, and content provenance. However, the dynamic nature of digital communication means that AI-generated text is often subject to subsequent modifications—either through benign human edits or deliberate tampering—making robust verification of content origin and integrity a pressing need.

Current watermarking approaches for LLMs primarily focus on global detection, answering whether a given text was generated by a watermarked model. These methods, however, are not equipped to handle the complexities introduced by post-generation edits. When text is altered after its initial creation—through insertions, deletions, or replacements—existing schemes struggle to reliably identify whether and where such modifications have occurred. This limitation undermines the utility of watermarking in real-world scenarios, where content may be collaboratively edited, paraphrased, or maliciously manipulated. As a result, there is a critical gap in the ability to localize edits and verify the integrity of AI-generated text, impeding efforts to ensure accountability, traceability, and trust in digital communications.

Technology

This technology provides a robust watermarking framework for large language models (LLMs), enabling both the embedding of hidden, identifiable patterns into generated text and the precise detection and localization of post-generation edits. The system works by partitioning the LLM’s vocabulary into distinct tagged subsets and cyclically biasing token selection during text generation, thereby embedding a deterministic combinatorial pattern. Detection is achieved through a sliding window approach that checks for conformity to the embedded pattern, while a per-token statistic is calculated to localize edits—such as replacements, insertions, or deletions—by identifying disruptions in the pattern. Notably, this method operates externally to the LLM, requiring no access to internal model parameters, and is lightweight enough for efficient integration into existing text generation pipelines.

What differentiates this solution is its ability to not only confirm whether a text is AI-generated but also to pinpoint the exact locations of post-generation modifications, a capability lacking in most existing watermarking methods. The combinatorial pattern-based approach is highly resilient, maintaining strong detection accuracy even for complex edit types and longer edit spans, and outperforms baseline watermarking schemes, especially in detecting deletions and short-span edits. The framework offers tunable trade-offs between watermark detectability and text quality, and its efficiency allows for real-time or batch processing at scale. This makes it particularly valuable for applications requiring content integrity verification, tamper detection, and provenance tracking in commercial LLM APIs, collaborative platforms, and compliance monitoring systems, setting a new standard for AI-generated content authentication.

Benefits

• Enables reliable embedding and detection of watermarks in AI-generated content (protein structures and text) without access to internal model parameters
• Accurately detects and localizes post-generation edits, including replacements, insertions, and deletions
• Preserves output quality and structural integrity while embedding watermarks, minimizing distortion
• Robust against common post-processing and adaptive attacks, ensuring watermark persistence
• Supports user identification and provenance tracking to protect intellectual property and enhance accountability
• Lightweight and efficient implementation suitable for integration with existing generative models and pipelines
• Applicable across domains including protein design for biosecurity and large language models for content authentication
• Offers scalable watermarking capacity with high recovery accuracy for practical deployment scenarios

Applications

• AI-generated content tamper detection
• Academic publishing integrity verification
• Compliance monitoring for LLM outputs
• Content auditing for enterprise platforms
• Authentication of public communications

Additional info

These systems introduce methods for embedding verifiable identifiers into AI-generated outputs. One framework subtly biases large language model token selection using cyclic combinatorial patterns to detect and localize text modifications. The other utilizes SE(3)-equivariant graph transformers and LoRA modules to embed binary codes into protein structures, enabling authentication and tracking while maintaining structural integrity.