Meta introduces TextSeal2, a new text watermarking algorithm that blends two secret keys into one virtual key with added noise to preserve signal integrity during speculative decoding. Experiments on Qwen3.5-9B show it achieves an 89% detection rate for 200-token outputs at a 0.1% false-positive rate, outperforming previous methods while maintaining generation diversity.
Text Watermarking
文本水印
Embedding invisible signals in generated text to identify its source
Curated Intelligence
As EU AI Act transparency rules take effect and major labs adopt text watermarking, concerns arise over degraded output quality. Research confirms that perfect distortion-free watermarking is technically impossible, with detection costs rising significantly for shorter texts.
Anthropic is expanding its imperceptible text watermarking to four additional Claude models, including Mythos 5 and Opus/Sonnet 4.x series. While newer versions enforce the mark by default, developers discovered that tweaking a single sampling parameter in the API allows bypassing the watermark on two older models. The initiative aims to identify AI-generated content without affecting meaning or readability, though it has sparked industry concerns regarding compliance and privacy.
OpenAI announced a phased approach to text watermarking to comply with the EU AI Act's machine-readable identification requirements. Global API users can now opt in for select models, while invisible watermarks will be added to eligible ChatGPT and Codex outputs in the EU within weeks. Acknowledging the technology's early-stage limitations, OpenAI emphasizes transparency and plans to expand detector access responsibly.
vLLM has officially integrated text watermarking to address the challenge of tracing AI-generated content. By influencing the generation process rather than perturbing discrete text, the feature embeds detectable signals while preserving the model's output distribution (non-distortion) and maintaining robustness against edits or truncation. This update offers developers an infrastructure-level solution balancing security with generation quality.
Google DeepMind and Hugging Face have officially integrated SynthID Text into Transformers v4.46.0, providing native watermarking for AI-generated text. The solution embeds invisible signals via modified sampling to help developers identify AI content and combat misinformation with minimal impact on output quality.