#

Key takeaways

• OpenAI published its approach to identifying generated text on October 5.

• The company says EU AI Act obligations motivate a phased deployment.

• API watermarking is off by default for eligible global customers.

• OpenAI plans invisible marks for eligible ChatGPT and Codex text in the EU.

• Detection is imperfect, especially for short or edited passages.

• No Moroccan rollout is announced in the source.

OpenAI's approach to text provenance

OpenAI published its approach to identifying generated text on October 5. The company says EU AI Act obligations motivate a phased deployment. It also says text detection remains imperfect.

The method, called textGrain, changes the statistical pattern of word choices. It does not add a visible label. OpenAI says this is meant to support provenance, not to make broad claims about authorship.

How the rollout works

Starting on the publication date, eligible API customers globally can opt in for selected models. API watermarking is off by default. OpenAI also plans to add invisible marks to eligible ChatGPT and Codex text generated in the European Union over the following weeks.

This is an EU rollout, not a global ChatGPT default. The source also says OpenAI is taking applications from researchers and expert organizations for limited detector access. It is not offering unrestricted public use.

What the detector can and cannot do

OpenAI says the detector is intended to assess whether a passage carries an OpenAI watermark. It is not meant to reveal a user, prompt, account, or conversation. That is an important limit on what the tool can tell readers.

The company also says short or constrained passages are harder to detect. Editing weakens the signal. In one example, replacing ten percent of words with synonyms reduced one measured rate from about ninety-two percent to sixty-six percent.

Reported evaluation results

OpenAI reports evaluation results at a one percent target false-positive rate. It identified about eighty percent of watermarked 200-token psychology passages. It identified roughly ninety-five percent of 400-token passages in the same evaluation.

These figures show that performance improves with longer text. They also show that detection is not perfect. The source makes clear that a watermark does not establish accuracy, ownership, legality, or the degree of human contribution.

Governance and operational considerations

The source emphasizes limits on interpretation. A failed detection does not prove human authorship. That matters because provenance tools can support review, but they cannot settle every question on their own.

OpenAI says its supported image and audio verification tools remain public. The text provenance rollout is separate from those tools. The source does not describe broader public access for the text detector.

Morocco relevance

The source reports no Morocco-specific rollout or program. For readers, the global lesson is simple: provenance tools can help, but their limits should shape how they are used.

What this means for readers

The main message is cautious deployment. OpenAI is pairing a phased rollout with limited access and explicit warnings about detection limits. That approach suggests the company sees provenance as useful, but not definitive.

For teams evaluating similar tools, the source points to three practical questions. How will the tool be used, who can access it, and what decisions will it support? Those questions matter because the detector cannot answer every authorship or legality issue on its own.