On 5 October OpenAI announced that it is starting to watermark ChatGPT text for EU users. Good news if you use ChatGPT. If your product calls GPT through the API, your text carries no watermark: in the API it is off by default, and you have to switch it on yourself. Few people know this, and in December details like this will decide whether you have anything to show.
Since 2 August, when the duty to tell people they are talking to AI came into force, model providers have moved further than in the previous two years. We checked the primary sources and sorted out what is in your hands.
What changed in two months
Since 2 August Anthropic has been watermarking text from supported Claude models with an invisible mark. New models are marked from day one, older ones are being added, 16 models so far, and the mark applies through the API as well. Images get C2PA metadata. The rollout on Amazon Bedrock completes on 17 October.
On 5 October OpenAI announced text watermarking: for ChatGPT and Codex users in the EU over the coming weeks, for API customers as an option right away. OpenAI has marked images since May. On the API default the company is explicit:
It will remain off by default in the API and will not become a global default at launch.
Google uses SynthID in the Gemini app and since August lets users remove the visible watermark; the invisible one stays. On Gemini API text, Google said "no" and then "yes" on its developer forum in August. The documentation still does not say.
Microsoft Azure points to Anthropic's marking for Claude models. Azure OpenAI GPT text is not marked yet; OpenAI promises it in the coming weeks.
Meta, Mistral and Cohere have signed the EU Code of Practice but have not published how they implement it. xAI has not signed.
On 24 September the European Commission published the official AI content icons: "Fully AI-Generated" and "Partially AI-Modified". Use is voluntary, binding for Code signatories.
What has not changed
The deadlines. Systems that were on the market before 2 August must have machine-readable marking by 2 December 2026 (Regulation 2026/1744, AI Act Article 111(4)). Code signatories have committed to interoperable detection tools by 2 February 2027. Fines go up to EUR 15 million or 3% of turnover.
Who marks what
Status on 9 October 2026, checked against provider documentation.
| Provider | Text | Images | Included in API | Text verification |
|---|---|---|---|---|
| Anthropic | yes, supported models | C2PA | yes | regulators and eligible organisations only |
| OpenAI | ChatGPT EU - in the coming weeks; API opt-in | C2PA + SynthID | only if switched on | approved researchers only |
| app yes; API - confirmed on the forum, not in the docs | SynthID + C2PA | unclear | waiting list | |
| Microsoft Azure | Claude - yes; GPT - not yet | C2PA | partly | - |
| Meta, Mistral, Cohere | not published | - | - | - |
| xAI | no | visible only | no | no |
What is not in the table: no provider publicly mentions marking Word, Excel or PDF files.
What this means if you build on these models
List the doors through which a model reaches you. The same Claude is marked through the direct API, through Bedrock from 17 October. The same GPT is marked through ChatGPT, through the API only if you have switched on "Text provenance".
Do not rely on what you cannot verify. Neither you nor your customer can check a text watermark - the detectors are for regulators and eligible organisations only. What you are left with is the provider's written confirmation. Ask for it.
Files are your responsibility. Document properties, a visible notice, PDFs with C2PA - nobody will do that for you. For PDFs the C2PA tooling has only been ready since late September.
And one correction to our August post: we thought one marking technique was enough in a closed business environment. The Code says otherwise. One technique is enough for free text (and text under 200 tokens needs no watermark at all); everything else needs two.
Where we stand
Chat answers our platform generates with supported Claude models carry Anthropic's watermark. For GPT models we switch "Text provenance" on. Marking of generated files (document properties, PDFs with C2PA) is being rolled out by 2 December.
If by December you have not written down which of your AI outputs are marked, with what and who confirms it, you have nothing to show. That is the whole job. Start with the list.
Facts checked on 9 October 2026.