Linear Ensembles Wash Away Watermarks: On the Fragility of Distributional Perturbations in LLMs
arXiv:2605.30501v1 Announce Type: new Abstract: Watermarking embeds statistical signatures in AI-generated text for detection and attribution. We reveal a fundamental vulnerability: when users access multiple models (today's reality), watermarks trivially fail. Watermarks perturb output distributions away from the original, and in competitive markets, these perturbations are typically independent across providers. We theoretically prove that averaging output probability distributions recovers the unwatermarked distribution with up to a second-order error term. Empirically, simply averaging 3-5
This research is emerging as AI-generated content proliferates and the need for attribution and provenance becomes critical, prompting explorations into both robust watermarking and its vulnerabilities.
A strategic reader should care because the fragility of watermarking affects trust, intellectual property, and regulatory efforts concerning AI-generated text, especially in multi-model environments.
The ability to attribute AI-generated text becomes significantly harder in a world where users interact with multiple LLM providers, potentially undermining current and future watermarking schemes.
- · Users employing multiple LLMs
- · Developers of unwatermarked open-source models
- · Those seeking to obscure AI text origins
- · LLM providers relying solely on watermarking for attribution
- · Regulators attempting to mandate AI content disclosure
- · Creators of original content
The primary effect is a reduced reliability of watermarks in practical, multi-model AI usage scenarios.
This reduction in reliability could lead to increased reliance on alternative or more complex methods for AI content detection and attribution, or a societal acceptance of ambiguity.
Ultimately, this might foster an environment where discerning human-generated from AI-generated text becomes a persistent and complex challenge, potentially impacting information integrity and legal frameworks.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.CL