What OpenAI Announced
On October 5, 2026, OpenAI announced it would begin watermarking ChatGPT-generated text for users in the European Union. The move is a direct response to the EU’s AI Act, which requires that AI-generated content be identifiable in a machine-readable format. The watermarking system, called textGrain, embeds an invisible statistical signal into the text by subtly altering word choices during generation — no hidden characters, extra spaces, or unusual punctuation.
Why This Is Happening Now
The EU AI Act (Regulation 2024/1689), specifically Article 50(2), mandates that generative AI providers make AI-generated text identifiable. The requirement took effect on August 2, 2026, with an extended deadline of December 2, 2026 for pre-existing systems like ChatGPT and Codex. OpenAI is a signatory to the EU Code of Practice published June 10, 2026, and the European Commission assessed it as adequate for demonstrating compliance.
How textGrain Works
textGrain works by slightly adjusting the probability distribution of words during text generation. A specialized detector can then check for this statistical pattern. According to OpenAI’s benchmarks:
- 400-token passages: ~92% detection rate
- 200-token passages: ~80% detection rate (at 1% false positive rate)
- After 10% synonym replacement: drops to ~66%
- After 25% synonym replacement: drops to ~17%
OpenAI states the watermark has no meaningful impact on text quality and that it “matched or exceeded” Google DeepMind’s SynthID for text in benchmark testing.
What This Means for ChatGPT Users
The watermarking applies to all ChatGPT plans — free through paid — for users in the EU. The rollout is phased over several weeks and is EU-only; it is not a global default. Users do not need to take any action — the watermark is applied automatically. Importantly, the watermark does not identify the specific user, account, prompt, or conversation. It simply indicates that the text was generated by an AI system.
API Customers Can Opt In
API customers worldwide can opt in to text watermarking for select models starting October 5, 2026. The feature is off by default in the API — customers must enable it through organization settings (Data controls → “Text provenance”) or project settings. Cloud partners including Microsoft Azure will offer watermarking for OpenAI model outputs accessed through their services.
Important Limitations
OpenAI is explicit about what the watermark cannot do:
- Does not guarantee reliable detection
- Does not verify accuracy or factual correctness
- Does not identify the specific user or account
- Does not determine who owns the text
- Does not measure human contribution
- Does not prove human authorship if no watermark is detected
- Does not prove the text was entirely AI-generated if a watermark IS detected
Detector access is restricted to approved researchers and expert organizations on a case-by-case basis. Cited partners include Cornell, ETH Zurich, and KInIT (Slovakia). OpenAI plans to open-source textGrain at an unspecified future date.
The Bottom Line
If you’re in the EU and use ChatGPT, your generated text now carries an invisible watermark. If you’re an API customer anywhere in the world, you can enable watermarking on select models. The system is imperfect — short passages, edits, paraphrasing, and translations can weaken or eliminate detection — but it represents a significant step toward AI content transparency under EU law.
Sources: OpenAI official announcement, EU Code of Practice, The Verge, Ars Technica
