Position: LLM Watermarking Should Align Stakeholders' Incentives for Practical Adoption

TL;DR

The paper argues that the slow real-world adoption of LLM watermarking stems from misaligned incentives among providers, platforms, and users. It analyzes three watermarking paradigms—model watermarking, LLM text watermarking, and in-context watermarking (ICW)—and examines how incentive structures shape their practicality, governance, and robustness in real settings. It advocates incentive-aligned, domain-specific designs, using ICW as a practical template where trusted parties control embedding and detection without compromising user experience. The authors propose design principles and directions for multi-bit provenance, threat-model driven robustness, and community engagement to enable faster, broader uptake while acknowledging regulatory roles and anti-detection market risks.

Abstract

Despite progress in watermarking algorithms for large language models (LLMs), real-world deployment remains limited. We argue that this gap stems from misaligned incentives among LLM providers, platforms, and end users, which manifest as four key barriers: competitive risk, detection-tool governance, robustness concerns and attribution issues. We revisit three classes of watermarking through this lens. \emph{Model watermarking} naturally aligns with LLM provider interests, yet faces new challenges in open-source ecosystems. \emph{LLM text watermarking} offers modest provider benefit when framed solely as an anti-misuse tool, but can gain traction in narrowly scoped settings such as dataset de-contamination or user-controlled provenance. \emph{In-context watermarking} (ICW) is tailored for trusted parties, such as conference organizers or educators, who embed hidden watermarking instructions into documents. If a dishonest reviewer or student submits this text to an LLM, the output carries a detectable watermark indicating misuse. This setup aligns incentives: users experience no quality loss, trusted parties gain a detection tool, and LLM providers remain neutral by simply following watermark instructions. We advocate for a broader exploration of incentive-aligned methods, with ICW as an example, in domains where trusted parties need reliable tools to detect misuse. More broadly, we distill design principles for incentive-aligned, domain-specific watermarking and outline future research directions. Our position is that the practical adoption of LLM watermarking requires aligning stakeholder incentives in targeted application domains and fostering active community engagement.

Figures (7)

• Figure 1: Example of model watermarking: an adversary fine-tunes, prunes, or illegally uses a protected model, and the LLM developers detect the unauthorized model.
• Figure 2: Incentive model for model watermarking among IP Owners, platforms, and users.
• Figure 3: Broken Incentive Model for LLM Text Watermarking: Users may switch to unwatermarked models, undermining both the LLM provider’s interests and the intended goal of reducing misuse.
• Figure 4: Illustration of two LLM text watermarking use cases. Left: Watermarking implemented by LLM Provider to detect self-generated data; Right: Watermarking implemented by the users to safeguard the user's document.
• Figure 5: Overview of In-Context Watermark.