Modern text watermarking changes the statistical choices a language model makes while generating tokens. The visible text remains ordinary text; the signal exists in the pattern of word or token choices across a passage.
A language model repeatedly chooses among many plausible next tokens. A watermarking scheme can subtly bias those choices according to a secret or controlled rule. Over enough text, the sequence forms a statistical pattern that a matching detector can test.
On October 5, 2026, OpenAI announced textGrain, a statistical text watermark intended for eligible ChatGPT and Codex output in the European Union, with rollout planned over the following weeks. API customers can opt in for supported models globally. OpenAI explicitly says textGrain does not insert hidden characters, spaces or watermark-only tokens; it changes statistical word-choice patterns.
Anthropic describes a similar probabilistic approach for future Claude models: nothing is added to the text, there are no hidden characters, and detection evaluates the pattern created by token selection. Anthropic also emphasizes that the signal says nothing about who used the system or who authored the final work.
Google DeepMind’s SynthID adjusts token probability scores during generation. Detection compares the resulting pattern with the expected behavior of watermarked and unwatermarked text. Google describes SynthID as a transparency signal rather than a universal solution for determining authorship.
Embedded during generation through token-selection behavior.
Invisible characters or unusual spacing inside a file. This is a different phenomenon.
Watermark detection generally benefits from longer passages because more token choices provide more statistical evidence. Editing, translation, constrained text and short excerpts can weaken detection. A missing watermark therefore does not prove human authorship, and a detected watermark does not establish who governed or approved the final document.
Primary sources: OpenAI — EU text provenance · Anthropic — Claude text watermark · Google DeepMind — SynthID