← Back to feed

OpenAI has launched textGrain, its text watermarking system, in direct response to EU AI Act mandates requiring machine-generated text to be identifiable in machine-readable form.

How textGrain Works

textGrain embeds an invisible statistical signal into a model’s word choices during generation — altering which tokens are selected at the margins to create a pattern the company’s detector can later recognize. It does not modify meaning and introduces no visible artifact.

OpenAI benchmarked textGrain against other approaches including SynthID for text and says it matched or exceeded competing systems on detection performance.

Three-Phase Rollout

The deployment proceeds in three steps:

1. Today, globally: API customers can opt in to text watermarking for select models. Off by default.
2. Coming weeks: All eligible ChatGPT and Codex EU text output will receive an invisible watermark automatically.
3. Restricted detector access: Applications are now open. Initial access is limited to approved researchers and expert organizations operating under the EU’s Code of Practice on AI transparency.

OpenAI says it will also release textGrain as open source, allowing third-party builders to embed and detect the signal independently.

Detection Performance

At a target false positive rate of 1%:

Text length	Detection rate
200-token passage (psychology)	~80%
400-token passage (psychology)	~95%
400-token passage (mathematics)	Substantially lower

Math and other constrained-vocabulary content produces weaker signals because there is less flexibility in word choice — fewer substitutions are available to encode the watermark.

Editing degrades the signal sharply:

Replacement rate	Detection rate (from ~92% baseline)
10% of words replaced	~66%
25% of words replaced	~17%

These limitations drove the decision to restrict detector access to approved organizations rather than opening it publicly. OpenAI notes that a detection result cannot measure human contribution, cannot verify the absence of a watermark in content not generated by OpenAI systems, and should not be treated as definitive proof of AI authorship.

No Quality Impact

OpenAI tested watermarked versus unwatermarked output across multiple benchmarks using GPT-6 Astra at maximum effort. Results were effectively flat:

• AA Intelligence Index: 49.57 vs 49.76
• AutomationBench: 34.09% vs 34.86%
• DeepSWE v1.1: 72.80% vs 71.68%
• Terminal-Bench 4.0: 53.90% vs 56.06%
• GPQA Diamond: 94.44% vs 93.94%

The variance across runs is larger than any consistent watermarking penalty.

EU Context

The EU AI Act requires generative AI providers to make AI-generated text identifiable in machine-readable form. OpenAI’s rollout puts it ahead of most competitors on direct compliance — Anthropic published its own Claude watermarking approach for text in August 2026 using a similar statistical method.

textGrain only applies to text. OpenAI’s existing Content Provenance API for images and audio remains separately available at openai.com/verify.