AI-generated text is starting to develop its own “invisible fingerprint.”
Starting November 7, SpaceXAI will enable text watermarking globally for several models in the Grok API, including Grok 4.7, 4.6, 4.5, and 4.3, as well as three versions of Grok 4.20: Reasoning, Non-Reasoning, and Multi-Agent.
The mechanism doesn’t rely on hidden characters or embedding information in text metadata. Instead, it adjusts token selection during generation, weaving the watermark signal into the text itself. It’s imperceptible to the human eye, but specialized detectors can identify it.
SpaceXAI said the move is in response to the EU’s Artificial Intelligence Act. Team administrators can also disable watermarking for specific models.
Andrew Kurlan noted that ChatGPT, in North America, is currently the only model that still doesn’t add invisible watermarks to generated text.
AI-generated content is entering a new phase: until now, users have mainly had to judge whether text sounds like it was written by AI. In the future, technical signals may make it possible to trace where it came from.
Of course, a watermark can offer clues about a text’s origin, but it cannot, on its own, prove who authored it.
Starting November 7, SpaceXAI will enable text watermarking globally for several models in the Grok API, including Grok 4.7, 4.6, 4.5, and 4.3, as well as three versions of Grok 4.20: Reasoning, Non-Reasoning, and Multi-Agent.
The mechanism doesn’t rely on hidden characters or embedding information in text metadata. Instead, it adjusts token selection during generation, weaving the watermark signal into the text itself. It’s imperceptible to the human eye, but specialized detectors can identify it.
SpaceXAI said the move is in response to the EU’s Artificial Intelligence Act. Team administrators can also disable watermarking for specific models.
Andrew Kurlan noted that ChatGPT, in North America, is currently the only model that still doesn’t add invisible watermarks to generated text.
AI-generated content is entering a new phase: until now, users have mainly had to judge whether text sounds like it was written by AI. In the future, technical signals may make it possible to trace where it came from.
Of course, a watermark can offer clues about a text’s origin, but it cannot, on its own, prove who authored it.