News

Regulation & Ethics

CONFIRMED: Grok flips its default invisible text watermark, exposing an industry-wide limit

xAI announced automatic text watermarking for Grok on November 7, and days later its own documentation said the opposite. The episode helps explain why no AI provider currently offers a public way to prove a text came from a model.

Ilustração editorial abstrata representando um sinal invisível embutido em texto digital, simbolizando marca d'água de IA

Watermark on, then off, then on again

On November 7, xAI announced that seven Grok models would start marking all text generated through its API with an invisible watermark by default. Days later, the company's own documentation was quietly updated to say the opposite: watermarking is off by default, and only team administrators can turn it on model by model. xAI never published a statement explaining the reversal, and the contradiction became visible to any developer who compared trade press coverage with the documentation page updated on November 9. Taken alone, the episode would look like a technical footnote. But it exposes something bigger: in under a month, the entire AI industry adopted a text tracking technology that most users will never be able to verify.

Three companies, three different policies

The race started with OpenAI, which announced textGrain on October 5, an invisible statistical signal embedded in the model's word choices. The company decided to apply it only to ChatGPT and Codex text inside the European Union, as a direct response to the EU AI Act, and left watermarking off by default in the API everywhere else. Anthropic went the opposite way and applies its watermark worldwide, not just in Europe. xAI chose SynthID Text, the technique Google DeepMind published in Nature in 2024, the same foundation Gemini uses. The result is an uneven map: the same text generated by three different assistants may or may not carry a traceable signal, depending only on each company's commercial policy at that moment, not on any single technical rule.

The detector almost nobody can use

The part that matters to everyday users is different: none of the three companies gives the public a way to check whether a text carries the mark. Access to OpenAI's detector is limited to approved researchers and organizations, case by case. xAI requires a formal request form. And even where access exists, OpenAI's own testing showed how fragile the signal is: swapping 10% of words for synonyms dropped detection from 92% to 66% in 400 token passages, and swapping 25% of words pushed it down to 17%. Shorter texts, under 200 tokens, are already harder to identify. In practice, this means any tool, paid or free, that claims to detect ChatGPT, Grok or Claude text for schools, companies or newsrooms isn't reading the official watermark, because it's secret and inaccessible. It's guessing from writing style, with all the error margin that implies. For anyone using AI to draft content, emails or customer service daily, the takeaway is direct: text watermarking today is a regulatory compliance tool between companies and regulators, not a way to publicly prove anything, for better or worse.

Sources

OpenAI: Our approach to EU text provenance rules; xAI docs: Output Watermarking; MACCUS: Grok will watermark API text from November 7; AI In Context: Grok API Watermarks Seven Models by Default.