Two of the world's leading AI labs are now invisibly marking the text their models produce. From November 7, 2026, text generated through the Grok API will carry a watermark by default, according to xAI's developer documentation. Four days earlier, on November 3, Anthropic extended Claude's text watermark to four more models. For founders building on these APIs, the change is quiet and consequential: the words your product generates may now carry a machine-readable signature you cannot see — and in Claude's case, cannot switch off.
What is happening, and when
The Grok date is documented in public. xAI's "Output Watermarking" page, updated October 31, 2026, states: "Starting November 7, 2026, responses from these models include a watermark. It's on by default, and you can opt out per model." The seven models covered are 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. The watermark applies to every text generation API — including the Responses API and the Chat Completions API — covering text responses, structured outputs, and tool-call arguments, on every API endpoint.
The Claude extension is less formally documented. Anthropic told API customers by email that from November 3, 2026, it would apply its text watermark to outputs from four older models — Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6 — as reported by The Decoder on October 31. Anthropic's public support table, checked on November 2, did not yet show the extension, so the November 3 date rests on that customer email rather than a public page. The four models join newer Claude models that have carried the watermark since August 2, 2026.
What a text watermark actually is
A text watermark is a statistical signature woven into the model's choice of words, not a hidden character or footnote. As a language model writes, it picks each word from a range of plausible options; watermarking subtly biases those picks toward a pattern a detector holding the right key can recognise later. The reader sees nothing unusual.
Claude's method is a variant of SynthID-Text, a technique Google DeepMind published in 2024. Anthropic says nothing is added to the text and there are no hidden characters. xAI says the same: watermarking "slightly shapes which tokens the model picks, so it adds no characters, markup, or hidden text," and does not change response quality or speed.
Unlike the Content Credentials metadata attached to Claude-generated files, which a free public checker can read, the text watermark has no public detector. For Claude, detection is a private preview for eligible organisations such as regulators, law enforcement, researchers, media fact-checkers, and educational organisations. Ordinary users cannot run the detector themselves.
Why now: the EU AI Act
Both companies give the same reason: compliance with the EU AI Act. Article 50 requires providers of AI systems that generate synthetic text, audio, images, or video to ensure outputs are marked in a machine-readable format and detectable as artificially generated. Anthropic signed the Article 50(2) Code of Practice on transparency of AI-generated content and applies its watermark worldwide on supported models. A December 2, 2026, deadline applies to providers whose systems were on the market before August 2.
As AI-generated text floods the internet, readers, platforms, and courts need a way to tell what a machine wrote. Watermarking is the industry's answer — less intrusive than a visible label, more durable than metadata stripped by copying and pasting.
Two designs, opposite postures
Although the driver is the same, the two implementations differ sharply. Claude's watermark cannot be switched off by the user, and there is no visible signal that a given piece of text carries it. Grok's watermark is on by default, but a team admin can opt out per model in the console — the change takes about five minutes and is audit-logged — and every API response carries an x-watermark: true header, so a builder can log whether output was marked.
There are also known gaps. Anthropic's own documentation says Opus 4.6 and Sonnet 4.6 may not reliably carry a detectable watermark when sampling parameters such as temperature are set to non-default values. And Grok's documentation leaves open whether the Grok apps, Grok Imagine, and the Google Cloud and Microsoft Foundry platforms that host Grok are covered.
What it means for founders
First, do not promise customers unmarked output. On a supported Claude model the text is watermarked and you have no detector to prove otherwise; on Grok, log the x-watermark header as your auditable record.
Second, treat the watermark as a compliance feature, not a bug. Customers in regulated industries, media, and education increasingly need to demonstrate content provenance; a product that shows its AI-generated text is properly marked has a selling point.
Third, expect convergence. ChatGPT and Gemini have watermarking schemes of their own, and machine-readable marking is becoming a default condition of using frontier models. Founders who design for it now — in their terms of service, their logging, and their user interfaces — will not have to retrofit later.
Angeled.com connects founders with investors.
Sources: xAI / SpaceXSI developer documentation, "Output Watermarking," updated October 31, 2026, as reported by Maccus and LLM Matchmaker, November 2026; The Decoder, October 31, 2026, as reported by Inoryva and Pasquale Pillitteri, November 2026; Anthropic support page, "How Claude marks AI-generated content," updated October 28, 2026.
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



