AI Text Watermarking: Which LLM APIs Mark Their Output

AI text watermarking, a hidden statistical pattern embedded through token selection and detectable only with corresponding keys, has become a significant compliance and provenance issue under the EU AI Act’s machine-readable marking requirements, which apply to new systems from August 2, 2026 and to pre-existing systems from December 2. Providers have adopted differing approaches: OpenAI offers opt-in textGrain watermarking for selected API models, Anthropic applies watermarking by default to supported Claude models, xAI plans default watermarking for specified Grok models, and Google has said Gemini API text carries SynthID, while coverage from providers such as Mistral and Meta remains undocumented. Whether a response is marked depends not only on its model name but also on its version, serving provider, account settings, generation configuration, region, and routing or fallback choices, meaning identical models can produce differently marked outputs through different platforms. Watermarks provide evidence that text passed through a particular system but cannot establish authorship, accuracy, ownership, or the absence of human editing, and their detection reliability falls with short passages, constrained generation, and revisions. Developers serving open-weight models can add watermarking through tools such as vLLM or Hugging Face Transformers’ SynthID-Text implementation, but must manage keys, detection, limitations, and documentation themselves. Applications using gateways such as Eden AI should pin or restrict verified provider routes, control failovers, and record the provider and configuration used for each response, since gateways do not create or verify watermarks.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.