Tuesday 6 October 2026
Morning EditionRSS
ZOTPAPER

News without the noise

ZOTPAPER
AI

OpenAI Introduces Text Watermarking, Takes Opt-In Approach Unlike Rival Anthropic

New textGrain system gives developers control over API output; EU users will see automatic watermarking

By LineZotpaper
Published
Read Time2 min
OpenAI has announced text watermarking for content generated through its API, using a new system called textGrain that embeds a statistical signal into AI output. Unlike Anthropic, which applies mandatory watermarking globally, OpenAI is making the feature opt-in for API customers, keeping it off by default. In response to the EU AI Act, the company will also begin automatically watermarking text from ChatGPT and Codex for users in the European Union in the coming weeks.

OpenAI's textGrain system works by subtly influencing the words the model chooses, favoring one suitable word over another when either would fit in a sentence. Over a long enough passage, these choices form a pattern that OpenAI's detector can identify.

“Starting today, API customers globally will be able to opt in to text watermarking for select models. Text watermarking will remain off by default in the API,” the company writes in a blog post. This lets customers decide how watermarking fits their transparency obligations and the experiences they provide to users. Developers can activate watermarking at the project or organization level, and once enabled, no changes to individual API requests are required.

The approach differs from that of Anthropic, which announced text watermarking for its Claude models in August. Anthropic applies the watermark globally to all supported Claude models, including those accessed through its API, and provides no equivalent opt-out for API developers. “Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from,” Anthropic has confirmed in its documentation.

OpenAI already uses other provenance technologies for generated content. Since 2024, it has added Content Credentials to generated images using the C2PA open standard, and it later added Google's SynthID watermarks to supported images.

§

Analysis

Why This Matters

  • AI-generated text is increasingly difficult to identify, raising concerns about misinformation, fraud, and academic dishonesty. Watermarking offers a technical tool to address this.
  • The EU AI Act sets new transparency requirements, making watermarking a compliance issue for companies operating in Europe. OpenAI’s automatic rollout for EU users is a direct response.
  • The differing approaches between OpenAI and Anthropic will test whether optional or mandatory watermarking is more effective in practice.

Background

Content provenance has been a priority for OpenAI since 2024, when it began adding C2PA credentials to images and later integrated Google’s SynthID. Text watermarking has been a tougher challenge because small changes to wording can preserve meaning while evading detection. The textGrain system addresses this by subtly steering word choices. The EU AI Act requires certain AI outputs to be marked, prompting the rollout for EU users of ChatGPT and Codex.

Key Perspectives

Developers: They gain flexibility to decide when watermarking fits their use cases, but an opt-in system may see low adoption if there is no regulatory push. Anthropic: By making watermarking mandatory and global, the company prioritises detection over developer choice, arguing that reliable provenance requires the feature on by default. Critics: Some worry that watermarking can be circumvented by editing or that competing standards could create confusion. The off-by-default design may also limit effectiveness.

What to Watch

  • Whether API developers enable watermarking at scale, and what adoption rates reveal about industry sentiment.
  • The EU’s enforcement of the AI Act and whether other jurisdictions introduce similar mandates.
  • How the two approaches compare over time: mandatory watermarking may be harder to evade, while optional watermarking may see more tailored use cases.

Sources

Zotpaper

Written by software from the reporting listed above, scored by an automated standards desk, and published without a person reading it first. If something here is wrong, tell the editor and it will be put right.

How we workReport an errorSubscribe