An Invisible Signature: OpenAI Watermarks ChatGPT and Codex Text in the EU — So What Should Enterprises Trust?

Telling AI-generated text apart from human writing has been an unsolved problem for years. On October 5, under regulatory pressure, OpenAI offered a concrete answer: an invisible text watermark called textGrain. Over the coming weeks it is being added to eligible text output across all ChatGPT and Codex plans in the EU.

How it works

• The watermark adds no hidden characters, invisible spaces or extra tokens; instead, as the model picks the next word, it gently shifts the probabilities of candidate words toward certain alternatives.
• The result is a pattern a reader cannot notice but a detector can pick up statistically.
• API customers worldwide can turn watermarking on for select models at the project or organization level; it is off by default.
• For now, only approved researchers and expert organizations can access the detector, case by case; private companies such as an HR department do not qualify.

Why now?

Article 50 of the EU AI Act requires providers of generative AI to make their output identifiable in a machine-readable way, and these transparency obligations have applied since August 2, 2026. OpenAI frames the watermark as its response to that requirement — which is why the rollout starts as an EU-only measure rather than a global default.

The limits: the trace is easy to erase

• In OpenAI's own tests, at a 1% false-positive target, detection is about 80% for 200-token responses and about 95% at 400 tokens.
• In a 400-token passage, swapping 10% of words for synonyms cuts detection from about 92% to 66%; swapping 25% drops it to 17%.
• Detection is unreliable on short, edited or translated text, and the company states plainly that a missing watermark does not prove a human wrote it.

Takeaways for enterprises

• Don't treat the watermark as an 'AI detector': access is restricted and results are weak on edited text. It should never be sole evidence in employee or candidate reviews.
• If you run customer-facing text generation in the EU, document how each model and provider you use complies with Article 50.
• Decide deliberately which API projects should have watermarking on; it is a product choice as much as a compliance one.
• Your strongest provenance evidence is your own records: knowing which model produced which text, when and from which prompt is far more reliable than a watermark.

Where remot fits

A watermark is about how text is recognized out in the world; inside the enterprise, the real question is being able to prove which model produced a given output. Because every model call flows through a single orchestration layer, remot logs each output with its model, provider, timestamp and policy context, and lets you manage settings such as which projects have watermarking enabled through central policies. That way, in an AI Act audit you rely on your own audit trail rather than on how robust a watermark is.