xAI (signed as SpaceXSI in its docs) has rewritten the "Output Watermarking" page of its API docs: the invisible text watermark that was due to turn on by default on November 7 is now documented as off by default, and team admins can turn it on per model in the Console for grok-4.7, 4.6, 4.5, 4.3, and the three grok-4.20 variants. The page also says the text watermark uses Google DeepMind's SynthID Text. Audio from the voice APIs (Text to Speech, Speech to Speech, custom voices) is watermarked with AudioSeal by default and can be opted out per model, but stays watermarked when the end user is in the EU, the EEA, or California. On November 14, Elon Musk posted on X that text watermarking with Grok is only on where required by law or requested by the user.
Key Takeaways
- ✓Text: off by default; team admins turn it on per model under Team settings → Model output watermarking in the Console, effective within 5 minutes and logged in the audit log (xAI docs)
- ✓Technique: the text watermark uses SynthID Text, published by Google DeepMind in Nature in 2024; it slightly shapes token choice and inserts no characters; the detection API is by application
- ✓Audio: Text to Speech, Speech to Speech, and custom voices are watermarked with AudioSeal by default and can be opted out per model, but stay watermarked when the request IP puts the user in the EU, the EEA, or California
- ✓A single request can force a watermark with the x-watermark: true header (watermark=true on WebSocket URLs); the x-watermark response header shows whether a response was watermarked
- ✓Musk said on X that Grok text watermarking is only on where required by law or requested by the user; xAI has not published the exact date of the change
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and pain point
On October 31, xAI published an "Output Watermarking" page in its API docs saying that from November 7 text output from grok-4.7 and six other models would carry an invisible watermark by default, citing Article 50 of the EU AI Act. Developers worried about their output being tracked, and opting out required an admin to confirm a legal exception. The same page now says text watermarking is off by default. On November 13 (UTC), the tech account Tech Dev Notes posted that API text watermarking was off by default, and on November 14 Elon Musk quoted it, saying Grok text watermarking is only on where required by law or requested by the user. xAI has not published the exact date of the change or a separate announcement.
Architecture
The docs now state that the text watermark uses SynthID Text, which Google DeepMind published in Nature in 2024: it slightly shapes token choice during generation and inserts no characters, and detection is statistical, with watermarked, not watermarked, and uncertain results. The control is still per team and per model, shared by every API key on the team. A new audio section uses AudioSeal, open-sourced by Meta, added before encoding so every output format carries it; teams can opt a voice model out, but audio stays watermarked when the request IP puts the end user in the EU, the EEA, or California.
Benchmarks
This is a platform policy change, so there are no benchmarks. xAI still has not published detection accuracy, false-positive rates, or how much text a reliable result needs, and there are no third-party tests yet.
Getting started
By default, Grok API text no longer carries a watermark, so no action is needed. If you need compliance marking, a team admin opens Team settings in the Console, finds Model output watermarking, checks the models, and saves; you can also log the x-watermark response header for audits. Teams using the voice APIs should note that audio served to users in the EU, the EEA, or California is always watermarked, and x-watermark: false does not remove it.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.