正文 · AI 翻译
IT Home reported on November 1 that SpaceXAI will begin adding invisible watermarks to text generated by Grok models on November 7. A detector identifies the signal to determine whether a passage of text carries a watermark invisible to the naked eye.
The company said the change is intended to meet the requirements of the EU AI Act. Watermarking has been enabled for the SpaceXAI API in all regions worldwide, including the grok-4.7, grok-4.6, grok-4.5, grok-4.3, grok-4.20 reasoning, grok-4.20 non-reasoning and grok-4.20 multi-agent models. Team administrators can choose to disable watermarking for a specific model.
IT Home noted that this hidden watermark does not rely on hidden characters or metadata; instead, it is incorporated into the process by which the model selects tokens when generating text.
AI governance observer Andrew Curran said that, at present, ChatGPT in North America is the only model that has yet to add invisible watermarks to text.