Unicode watermarking methods tested against LLMs, with mixed results

A new paper from arXiv analyzes the security and detectability of Unicode text watermarking methods against various large language models. Researchers tested ten watermarking techniques across six models, including GPT-5, GPT-4o, Llama 3.3, and Claude Sonnet 4. The study found that while advanced reasoning models can detect watermarked text, they generally fail to extract the watermark without access to the source code. AI

IMPACT This research highlights potential vulnerabilities in text watermarking against advanced LLMs, impacting data security and content authenticity.

RANK_REASON Academic paper published on arXiv detailing security analysis of watermarking methods against LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Claude Sonnet 4
• Gemini 2.5 Pro
• GPT-4o
• GPT-5
• large language models
• Llama 3.3
• Malte Hellmeier
• Teuken 7B
• Unicode

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →