Grok API Text Watermarking Explained: Is It On, and How to Check Your Responses

The Grok API was set to watermark text by default from 7 November, then xAI’s docs said it is off. Here is what happened and how to check the x-watermark header yourself.

Reviewed by Mindloop Editorial on . Every link in the Sources section at the end was re-checked that day.

Graphic of a Grok API response with an x-watermark header toggled between true and off.

Imagine your AI provider emails customers to say every sentence its models write will carry an invisible signature from 7 November. Then, days after that date, the documentation says the signature is off by default. That is the story of text watermarking in the Grok API, and it comes with a header you can check yourself.

Key takeaways

  • A customer email said Grok API text watermarking would be on by default from 7 November; the docs now say off.
  • The watermark is invisible: SynthID Text slightly shapes which tokens the model picks, adding no characters or markup.
  • Check any response yourself: an x-watermark: true header means watermarked, anything else or no header means not.
  • Team admins switch it on per model in the Console, and changes apply to new requests within 5 minutes.

This guide explains what happened, how the watermark works, and how to see whether your own Grok API responses carry it. All details are from xAI’s documentation and press coverage as of 15 November 2026. The request and Python examples below follow xAI’s docs. We syntax-checked them (the JSON body parses and the Python compiles); live requests need your own paid API key, so any sample output shown is the docs’ own. (The docs and Console now call the company SpaceXSI. Some coverage says xAI.)

On, then off: what happened

On 31 October, reports described a customer email saying watermarking “is on by default for all teams and will be enabled on November 7, 2026”, according to FourWeekMBA, which cited a screenshot posted on X. Crypto Briefing reported the same plan. Seven models were named: grok-4.7, grok-4.6, grok-4.5, grok-4.3 and three grok-4.20 variants.

Then the documentation changed. TeslaNorth reported on 11 November that the Output Watermarking page, last updated 10 November, now tells admins to enable the feature per model. Today the docs state plainly: “Text watermarking is off by default.”

FourWeekMBA relays a report from The New Stack that the change happened overnight between 9 and 10 November, with no public explanation. FourWeekMBA also quotes Elon Musk on X: “Text watermarking with Grok is only on where required by law or requested by the user.” We could not read The New Stack directly, so treat that detail as second-hand.

Did you know?

Grok voice audio is watermarked by default with AudioSeal, and it stays on for end users in the EU, the EEA or California even if a team opts out.

Why a watermark at all

The docs say watermarking “can help enable compliance with the EU AI Act”. FourWeekMBA points to Article 50(2), which it quotes as requiring AI outputs to be “marked in a machine-readable format and detectable as artificially generated or manipulated”. It also reports a 2 December 2026 compliance date for systems already on the market before 2 August 2026. That is FourWeekMBA’s reading of the law, not xAI’s. It adds that OpenAI announced API text watermarking on 5 October 2026, off by default.

By the numbers

5 minutesfor setting changes to apply

2 December 2026reported EU compliance date

2024SynthID Nature paper published

How the watermark works

The watermark is invisible. According to the docs, it slightly shapes which tokens (word-sized chunks) the model picks as it writes, so it adds no characters, markup or hidden text. xAI says it uses SynthID Text, the technique Google DeepMind published in a 2024 Nature paper.

Detection is statistical. A detector scores a piece of text and reports watermarked, not watermarked, or uncertain, and longer text gives a stronger signal. TeslaNorth, describing DeepMind’s own limits, says heavy rewriting or translation can lower confidence. So it works as one identification signal with real limits.

xAI says watermarking does not change quality and has a negligible effect on speed. That is the vendor’s claim.

Voice is different. Audio from the voice APIs is watermarked by default using AudioSeal, and the docs say it stays on for end users in the EU, the EEA or California even if your team opts out.

Video: SynthID – Watermarking and identifying AI-generated text (Google for Developers)

What you need

  • An xAI API key, exported as XAI_API_KEY.
  • A supported text model, such as grok-4.7.
  • A team admin on your account, if you want to change the setting.

Step 1: Check whether a response was watermarked

Read the x-watermark header on the Responses or Chat Completions reply. The -i flag prints headers:

Terminal
curl -i https://api.x.ai/v1/responses \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-4.7",
    "input": "Write a haiku about the ocean."
  }'

In Python, the docs show the OpenAI SDK with raw response access:

Code
import os
from openai import OpenAI

client = OpenAI(
    api_key=os.getenv("XAI_API_KEY"),
    base_url="https://api.x.ai/v1",
)

raw = client.responses.with_raw_response.create(
    model="grok-4.7",
    input="Write a haiku about the ocean.",
)
response = raw.parse()

print(raw.headers.get("x-watermark"))
print(response.output_text)

x-watermark: true means that output was watermarked. Any other value, or no header at all, means it was not.

Step 2: Turn text watermarking on (admins only)

The docs list these Console steps:

  1. Open Team settings and find the “Model output watermarking” row. It shows On (all models), Off (all models) or Custom.
  2. Select the settings icon, then tick each model you want. “Select all” flips every model at once.
  3. Select Save. If you turned a model off, confirm with “Opt Out”.

The setting covers every API key on the team and applies to new requests within 5 minutes. The docs describe this Console toggle as the way to change it, and say no code changes are needed once it is on. Every opt-in and opt-out lands in the team audit log.

Four-step flow for enabling Grok API text watermarking in Team settings, saving, and re-checking the header

Step 3: Re-run the check

Wait five minutes, then repeat Step 1. If the model is switched on, you should see x-watermark: true.

Step 4: Ask for detection access

Detection is granted by application. The docs link a request form for the detection API.

Common errors and fixes

  • The header is missing. That counts as “not watermarked”. Check the setting for that exact model, and wait five minutes after saving.
  • You cannot see the watermarking row. Only team admins see it. Ask yours.
  • A model has no checkbox. The docs say some models are locked on for a team, and locked models cannot be turned off in the Console.
  • Opting out asks for confirmation. Turning a model off means confirming that a legal exception applies and that you are solely responsible for any marking or disclosure rules.
  • Detection says “uncertain”. Docs and press both say short text gives a weak signal. Try a longer sample.
  • Audio is watermarked despite x-watermark: false. For voice, that header does not override a location requirement. Any value other than true or false returns a 400 error.
Comparison of Grok text and voice watermarking defaults, techniques and opt-out rules side by side

What this means for you

If you build on Grok’s API, the practical move is to check the header on your own responses instead of trusting last week’s news. Defaults have already flipped once. Our guide to adding Grok to Zoom, Google Meet or Teams calls covers a different Grok product. If you work across providers, our Gemini 4 Hydrogen API guide covers Google’s side.

Frequently asked questions

Is Grok API text watermarking on by default?

No. A customer email said it would be enabled on 7 November 2026, but the xAI documentation, last updated 10 November, now states that text watermarking is off by default and admins enable it per model.

How do I know if a Grok API response is watermarked?

Read the x-watermark header on the Responses or Chat Completions reply, for example with curl -i. A value of true means the output was watermarked; any other value, or no header at all, means it was not.

Does watermarking change Grok’s output quality or speed?

xAI says watermarking does not change quality and has a negligible effect on speed. That is the vendor’s claim. The watermark adds no characters, markup or hidden text.

Can a SynthID text watermark be removed?

Detection is statistical and can weaken. TeslaNorth, describing DeepMind’s own limits, says heavy rewriting or translation can lower confidence, and short text gives a weak signal, so it works as one signal with real limits.

Quick quiz

Which technique does xAI say it uses for text watermarking?

  • AudioSeal
  • Content Credentials
  • SynthID Text

xAI says it uses SynthID Text, the technique Google DeepMind published in a 2024 Nature paper.

Who can change the watermarking setting in the Console?

  • Team admins
  • Any API key holder
  • Only xAI support

Only team admins see the Model output watermarking row in Team settings.

What does detection report for a short piece of text?

  • Always watermarked
  • Often uncertain
  • A 400 error

Docs and press both say short text gives a weak signal, so detection may say uncertain.

Sources

Links last checked on 15 November 2026.

Written by Mindloop Editorial

The Mindloop Experience editorial team writes clear, carefully sourced explainers and how-tos on AI and technology, software and the business of tech, for anyone who likes to know how things actually work.

Meet the editorial team

Responses

Write a response

More from Mindloop Experience

See all from Mindloop Experience

You can clap up to 50 times.