At each generation step, a model picks among several near-equally good next tokens; watermarking schemes such as Google DeepMind's SynthID-Text (Nature, 2024), which builds on a 2022 proposal by Scott Aaronson, use a secret key to bias that low-stakes choice instead of a plain random number generator. The resulting token sequence carries a detectable statistical skew over a long enough passage, while remaining indistinguishable to a human reader and costing no extra tokens or latency. Anthropic adopted a version of SynthID-Text for Claude in August 2026 to comply with the EU AI Act's Code of Practice on Transparency of AI-Generated Content. The signal is sparse on factual or code-heavy text, where fewer word choices are equally valid, and it degrades or disappears after heavy paraphrasing, translation, or a near-total rewrite.