Skip to content
GrokDetect
EN

Can SynthID detect Grok text?

By Yang Zheng · Updated

Not with Google's public tools. As of 2026-11-12, SpaceXSI says Grok's API text watermark uses SynthID Text, but Google's SynthID Detector checks only images, video and audio, and its partner list names OpenAI, NVIDIA, Kakao and, soon, Apple. Grok text can be verified only by SpaceXSI, through a detection API it grants by application. For everyone else, the open option is a style check such as GrokDetect.

What SpaceXSI says

SpaceXSI's watermarking page, last updated 11 November 2026, says: "Our implementation uses SynthID Text, the text watermarking technique published by Google DeepMind in a Nature paper in 2024." Detection is statistical and returns watermarked, not watermarked or uncertain, and longer text gives a stronger signal. Access to detection is "granted by application."

The same page says text watermarking is off by default, so most Grok text carries no mark. The model list, the Console setting and the response header are covered on the Grok watermark status page and are not repeated here.

What Google's tools check

SpaceXSI's application-based API reads like the semi-private option, though its page does not use that term. Nothing on either company's pages says a Google tool can verify Grok text.

Who can check what

Route What it tells you Who can use it
x-watermark response header Whether that one API response was marked The developer who made the call
SpaceXSI detection API Watermarked, not watermarked or uncertain Approved applicants only
Google SynthID Detector Images, video and audio; text and xAI are not listed Anyone, but not for Grok text
Style check such as GrokDetect How closely writing matches Grok and other model families Anyone, from 100 words

What edits do to a text watermark

A watermark lives in the word choices of the original output, so changing the words changes the evidence. Google's guidance says SynthID is robust to "modifying a few words, or mild paraphrasing," while "detector confidence scores can be greatly reduced when an AI-generated text is thoroughly rewritten," and confidence can also drop after translation. It adds that the mark is less effective on factual responses.

The 2024 Nature paper is similarly plain: generative watermarks are weakened by edits, longer texts carry more evidence, and "no text detection method is foolproof." SpaceXSI's page does not state how its own marks behave after editing.

The practical result is that a present mark is useful evidence, an absent mark says almost nothing, and most Grok text starts without one.

Where a style check fits

A style check does not look for a mark. GrokDetect, which runs on Veread, Fileverity's text model, reads how a text is written and reports an AI verdict, the passages that carry the signal and a similarity to Grok and other families. On the facts page, 0.5% of human-written texts in seven languages were wrongly flagged, and 95.5% of AI-written and AI-assisted texts of every kind were detected. That last figure includes AI-assisted text, but a heavy human rewrite can still lower the signal.

The two methods answer different questions. A watermark says a specific vendor's system produced a text. A style check says the writing looks machine-made and which family it resembles.

Check a text now

Paste at least 100 words into GrokDetect for a free check. The result is writing-style similarity, not proof of authorship, and it does not read SpaceXSI's watermark. For a post thread from X, join the replies first, as described in checking X posts for Grok.

Sources: SpaceXSI, "Output Watermarking," last updated 11 November 2026; Google DeepMind, "SynthID"; Google AI for Developers, "SynthID: tools for watermarking and detecting LLM-generated text"; Dathathri et al., Nature, 23 October 2024.

Questions and answers

Can Google's SynthID Detector check Grok text?

No, as of 12 November 2026. Google's page says the SynthID Detector checks images, video and audio, and names OpenAI, NVIDIA, Kakao and, soon, Apple as partners. It lists neither text nor xAI.

Does Grok really use SynthID?

SpaceXSI's documentation says its implementation uses SynthID Text, the technique Google DeepMind published in Nature in 2024. Text watermarking is off by default and has to be switched on per model by a team admin.

Who can verify a Grok text watermark?

Only SpaceXSI, through a detection API it grants by application. Its x-watermark response header also tells the person who made an API call whether that response was marked.

Does editing a text remove the watermark?

It can weaken it. Google says SynthID is robust to a few changed words and mild paraphrasing, but detector confidence can be greatly reduced when text is thoroughly rewritten or translated.

What can I use if there is no watermark to read?

A style-based check such as GrokDetect, which needs at least 100 words and reports similarity to Grok's writing style, not proof of authorship.