On 14 November 2026, Elon Musk wrote: "Text watermarking with Grok is only on where required by law or requested by the user." Official documentation for the Grok API, published at docs.x.ai and last updated on 13 November 2026, agrees: "Text watermarking is off by default." Team administrators can turn it on, model by model. If they do not, the generated text carries no hidden mark.
A hidden watermark is not a substitute for a workplace policy on when staff should disclose AI-written work, and it is not a substitute for training.
What actually changed
The watermark in the official docs is invisible. It adds no label, footer or hidden characters. It slightly shapes which words the model chooses, so a detection tool can later try to tell whether the text was marked. Detection is statistical: the result may be watermarked, not watermarked, or uncertain, and longer text gives a stronger signal. Access to detection is granted by application, not automatically.
This is the API setting, not a claim about every Grok product or every tool your staff might use. Audio is treated differently: voice API output is watermarked by default, and that mark stays on when the end user is in the EU, the EEA or California.
Musk's wording also goes further than the documentation. The docs do not say text watermarking switches on automatically wherever the law requires it. They say administrators turn it on, and that turning a model off means confirming an exception under applicable law and taking responsibility for any marking or disclosure duty. If your organisation builds on the API, read the current docs and take advice rather than assuming the provider will mark outputs for you.
How the main AI providers compare
Official documentation checked on 16 November 2026 shows that providers do not treat text watermarking the same way. Grok API text marks are off by default. Claude applies them on supported models worldwide. OpenAI applies them to eligible ChatGPT text in the EU and leaves them opt-in in the API. Google describes SynthID text marks on the Gemini app and web. Microsoft and Meta have not published a default text watermark for their main chat products. Image, audio and video marks, and who can use a detector, also differ.
Scroll sideways to see all five columns.
| Provider | Text | Images | Audio and video | Public detector |
|---|---|---|---|---|
| xAI (Grok API) | Off by default | Not published | Audio: on by default. Video: not published | No. Text and audio detection by application |
| OpenAI (ChatGPT / API) | On by default for eligible ChatGPT text in the EU. Opt-in in the API | On by default (C2PA and SynthID) | Audio: on by default. Video: no current video product (Sora discontinued) | Yes, for images, audio and PDFs. Text detection by application |
| Google (Gemini) | On by default in the Gemini app and web. API / Vertex: not published | On by default (SynthID, and C2PA where supported) | On by default for supported audio and video | Yes, for images, video and audio. Text: not published as a public tool |
| Anthropic (Claude) | On by default on supported models, worldwide | On by default (C2PA on supported files) | Not published | Yes, for file Content Credentials. Text detection in private preview |
| Microsoft (Copilot / Azure OpenAI) | Not published | On by default for supported Azure images (C2PA). M365 visible mark is opt-in | Not published. Foundry lists audio provenance as available but does not state a default. M365 visible or spoken marks are opt-in | Yes, for images, audio and video. Text: not published |
| Meta (Meta AI) | Not published | On by default (Content Seal, plus earlier visible "Imagined with AI" labels) | Not published | Yes, at meta.ai/identification, for images, video and audio (Meta still calls it a preview). |
The law is narrower than the headline
In the UK, there is still no legislation that requires AI-generated text to be labelled or watermarked. The House of Commons Library said so in a briefing published on 20 January 2026. On 23 October 2026, the government said it is considering whether further action is needed, including a taskforce on labelling AI-generated content, due to launch by the end of the year. That is a policy process, not a current legal duty on employers.
The EU is in a different position. Article 50 of the EU AI Act has applied since 2 August 2026. The European Commission says providers of generative systems must apply a machine-readable mark to synthetic text, images, video and audio so the output can be detected as AI-generated. For systems already on the market before 2 August 2026, that duty has a transition period until 2 December 2026. Deployers (organisations using the system in a professional capacity) must clearly label deepfakes, and must label AI-generated or manipulated text published to inform the public on matters of public interest, unless that text has had proper human review or editorial control. Superficial spelling or grammar checks do not count.
Most internal workplace writing in a UK organisation will not be a public-interest publication in the EU, though some of it might be if you publish to EU audiences. This is not legal advice: ask counsel how Article 50 applies to your facts. The practical point is simpler. A legal mark is aimed at machine detection. It will not tell a manager whether a colleague used AI on a policy, a job advert or a board paper.
Detection will not run your policy for you
Employers cannot treat watermark detection as a safety net. Staff already use a mix of approved and unofficial tools. A detector built for one provider's mark will not see text from a tool that never applied one. Short passages may come back as uncertain even when a mark is present. A statistical score is not the same as a conversation with the person who did the work.
The risk is not simply that AI-written work goes unnoticed. It is that people hide the use, or avoid useful tools, because they do not know what they are allowed to do. That is the same pattern we wrote about when Imperial's research on shadow AI found younger staff in particular using AI without telling their employer.
A workable position has three parts: a written rule people can follow, a sanctioned way to do the work, and training so staff can apply the rule in real tasks. The rule should say when AI use must be disclosed, which tools are approved, what must not be pasted into a tool, and who checks the work when a mistake would be costly or public. Human review still matters. If someone has examined the substance of the work, and is accountable for it, you have a standard of quality that no hidden mark can provide.
What employers can do next
The AI & Automation Practitioner Level 4 apprenticeship (ST1512) is the levy or government-funded route for people who will turn everyday processes into approved, documented workflows. It is live online, needs no coding, and takes around 15 months. Early modules include a working acceptable-use position. Later modules cover process discovery, human-in-the-loop automations, and the guardrails that keep that work safe.
For senior leaders who need to set the organisation's position, the AI for Leaders workshop is the shorter route. It is a paid leadership programme, not an apprenticeship, typically three days across three to six weeks. It covers AI strategy, risk, talent and a governance position appropriate to the sector, which is where a disclosure policy belongs.
- Next AI & Automation Practitioner Level 4 cohorts: 14 December 2026 and 25 January 2027, live online.
TESS Group has been delivering apprenticeships since 2009. Across our apprenticeships, 72% of apprentices achieve a distinction (results from 2024 to 2026), and we're rated 4.9/5 from 793+ reviews (Google and Trustpilot combined).
Find out more about the AI & Automation Practitioner Level 4 or book a complimentary 30-minute discovery call to talk through a disclosure policy, a Level 4 cohort, or the AI for Leaders workshop.
Sources: Elon Musk, 14 November 2026, https://x.com/elonmusk/status/2121644169462681818; Tech Dev Notes (@techdevnotes), 13 November 2026, https://x.com/techdevnotes/status/2121370786208375136; Mark Kretschmann (@mark_k), 15 November 2026, https://x.com/mark_k/status/2121975392554619298; official Grok API documentation, "Output Watermarking", https://docs.x.ai/developers/advanced-api-usage/watermarking, last updated 13 November 2026; European Commission, "Transparency obligations under Article 50 of the AI Act", https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act, and "Quick Facts: Transparency rules for AI systems", https://digital-strategy.ec.europa.eu/en/factpages/quick-facts-transparency-rules-ai-systems; EU AI Act Service Desk, implementation timeline, https://ai-act-service-desk.ec.europa.eu/en/ai-act/eu-ai-act-implementation-timeline; House of Commons Library, "AI content labelling", 20 January 2026, https://commonslibrary.parliament.uk/research-briefings/cbp-10467/; UK Parliament written answer UIN 35904, 23 October 2026, https://questions-statements.parliament.uk/written-questions/detail/2026-10-19/35904; OpenAI, "Provenance signals in OpenAI-generated content", https://help.openai.com/en/articles/8912793-provenance-signals-in-openai-generated-content, and https://openai.com/index/eu-text-provenance/; Google DeepMind, SynthID, https://deepmind.google/models/synthid/, and Gemini Help, https://support.google.com/gemini/answer/16722517?hl=en; Anthropic, https://platform.claude.com/docs/en/build-with-claude/text-watermark and https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content; Microsoft Learn, https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/content-safety/provenance-disclosure and https://learn.microsoft.com/en-us/copilot/manage/watermarks; Meta, https://ai.meta.com/blog/introducing-muse-image-muse-video-msl/; Meta AI identification tool, https://meta.ai/identification.