SynthID Text watermarking added to Transformers v4.46.0 for AI-generated content detection

AI Impact Summary

SynthID Text introduces a watermarking scheme for AI-generated content by integrating a g-function-based watermark into Transformers generation and providing a detector classifier. This enables governance and trust programs by allowing verification of authorship for text produced with any compatible LLM, with detectors shareable across models that use the same tokenizer and deployable via private HF Hub. Practical considerations include per-model watermark configuration (keys and ngram_len), a detector training pipeline (minimum ~10k examples), and known limitations such as robustness to paraphrasing or translations and potential impact on generation quality.