OpenAI rolls out optional text watermarking for API, keeping it off by default

OpenAI has launched an opt-in watermarking system for text generated through its API, marking a significant step in proving the provenance of AI-written content. Developers can now enable the technology, which the company calls textGrain, on supported models starting immediately. Notably, watermarking remains off by default—a deliberate contrast to how some competitors handle the feature.

The textGrain approach works by embedding a detectable pattern into generated text without substantially altering its quality. Rather than inserting visible markers, the system subtly nudges a model's word choices, favoring one suitable option over another when multiple would work equally well. Over longer passages, these preference patterns form a statistical signal that OpenAI's detector can identify.

How textGrain differs from competitors#

Anthropic, which rolled out text watermarking for Claude in August, took a different route by applying the watermark globally across all supported models and API access. The company explained that regional limitations weren't technically feasible at launch, and offers no equivalent opt-out for developers. OpenAI's decision to make watermarking voluntary gives developers more granular control—they can activate the feature at the project or organization level and choose which models to apply it to.

OpenAI is also developing its own technology rather than adopting existing solutions like Google's SynthID or Meta's TextSeal. The company says textGrain offers superior control over the tradeoff between watermark detectability and response diversity, and claims it matched or exceeded SynthID's performance in internal testing. OpenAI plans to open-source the method to encourage broader adoption and refinement.

Detecting the watermark: where it works—and doesn't#

The watermarking approach has meaningful limitations. OpenAI's testing shows detection rates vary significantly depending on text length and type:

• Longer passages perform better: The detector catches roughly 80% of watermarked 200-token passages and around 95% of 400-token passages in domains like psychology (at a 1% false-positive rate)
• Constrained domains are harder: Heavily structured content such as mathematical writing shows lower detection rates
• Code presents a special challenge: Since programming languages offer fewer plausible word choices than prose, watermarking code is substantially more difficult

Editing poses another obstacle. When users replace 10% of words in a 400-token passage with synonyms, detection drops from roughly 92% to 66%. Replacing a quarter of the words collapses detection to just 17%.