> ## Documentation Index
> Fetch the complete documentation index at: https://docs.praxis-ai.com/llms.txt
> Use this file to discover all available pages before exploring further.
# AI-Generated Content Transparency
> How Praxis AI and its model providers disclose AI interaction and mark AI-generated content under the EU AI Act (Article 50).
## Overview
Since **2 August 2026**, Article 50 of the EU Artificial Intelligence Act requires generative AI to be transparent about what it is. Three obligations matter for a Digital Twin:
| Obligation | What the Act requires | How it is met on Pria |
| ----------------------------------------- | ---------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **AI disclosure** — Art. 50(1) | People must be told they are interacting with an AI system, unless it is obvious. | Every Digital Twin is presented as an AI-powered expert, and the conversation interface reminds users they are talking to an AI. See [AI disclosure in the interface](#ai-disclosure-in-the-interface). |
| **Machine-readable marking** — Art. 50(2) | Generated text, images, audio and video must carry a machine-readable mark that identifies them as AI-generated. | The mark is applied **inside the model** by each AI provider. Praxis AI relays outputs unchanged, so the mark reaches your conversations and your vault. See [How marking reaches your Digital Twin](#how-marking-reaches-your-digital-twin). |
| **Labelling** — Art. 50(4) | Deployers must clearly label deepfakes, and AI-generated text published on matters of public interest. | Your responsibility as the institution deploying the twin. See [What your institution should do](#what-your-institution-should-do). |
Systems already on the market before 2 August 2026 have until **2 December 2026** to complete machine-readable marking. Praxis AI's main model providers — Anthropic, OpenAI, Google, Microsoft and Mistral among them — signed the European Commission's *Code of Practice on Transparency of AI-Generated Content* in July 2026 and are rolling the marking out across their models ahead of that date.
This page describes how the platform behaves as of **September 2026** and is updated as providers extend their marking. It is not legal advice: your own obligations depend on where and how you deploy your Digital Twin.
## How marking reaches your Digital Twin
Praxis AI is a multi-model platform: every reply is produced by a model from one of the providers in the [platform catalog](/mdx/admin-guide/platform-models). Content marking is applied **by the provider, inside the model**, so it is already in the output before it reaches Praxis AI — on every channel: web, mobile, Canvas and other LMS launches, the Web SDK, embedded avatars and the REST API.
Praxis AI's part is to keep those marks intact:
* **Text is relayed word for word.** Pria streams the model's words to you exactly as they are generated. What Pria adds around a reply — citations, the Search Results row, the Memory row, Tool Details — is attached to the message, not written into the model's sentences. Copying, downloading, exporting or sharing a reply carries the mark with the text.
* **Generated images and videos are stored as delivered.** Image and video generation results are saved to your vault byte for byte as the provider returned them, so the invisible watermark and any Content Credentials (C2PA metadata) they carry stay with the file.
* **Praxis AI never removes, alters or hides a provider's mark**, and adds no mark of its own.
* **Nothing changes for you.** The marks are imperceptible: a text watermark adds no characters, tokens or metadata, does not change the meaning, quality or readability of a reply, and has no effect on credits, latency or API formats. It contains no information about you, your institution or your conversations. There is nothing to configure and nothing to turn on.
**Bring Your Own Model.** A [custom model endpoint](/mdx/admin-guide/ai-models#bring-your-own-ai-model-byom) carries whatever marking its own provider applies. Praxis AI relays its output unchanged in the same way, but cannot add a mark the model does not produce.
## AI disclosure in the interface
Praxis AI presents every Digital Twin as what it is: an AI-powered version of an expert, trained on that expert's knowledge and style. In the 2026 interface a standing reminder sits under every conversation — *"\ can make mistakes — verify important info."* — and Digital Twins, Assistants and voice avatars are described as AI throughout the product and this documentation.
When you embed a twin in your own page or LMS course, keep that context visible: introduce it as an AI Digital Twin, and do not present it as a person.
## Provider status
Which marks your twin's output carries depends on the models selected under [Personalization](/mdx/admin-guide/personalization) for each Model Use. Status as of September 2026:
| Provider | Text | Images, audio, video | Code of Practice |
| ---------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------- |
| **Anthropic** (Claude — direct and through Amazon Bedrock) | **Text watermark.** Claude Fable 5.1 and every model released after 2 August 2026 carry it from launch. **Claude Opus 5 from 9 September 2026**; remaining Claude models over the following weeks. Applied at the model layer, so it is present on every platform where Claude is served, including Amazon Bedrock. | Content Credentials (C2PA) on generated images. | Signatory |
| **OpenAI** (GPT, image and audio models) | Not yet. OpenAI has stated its goal to extend provenance signals to text as standards and tooling mature. | Content Credentials (C2PA) plus a SynthID watermark on generated images (since May 2026) and audio (since July 2026), across ChatGPT, the API and Codex. | Signatory |
| **Google** (Gemini) | SynthID text watermark on Gemini output. | SynthID on generated images, audio and video. | Signatory |
| **Mistral AI** | Signatory to the Code's provider section; per-model marking status is published by Mistral. | — | Signatory |
| **Amazon** (Nova models on Bedrock) | — | Invisible watermark plus Content Credentials on Nova Canvas images and Nova Reel videos. Claude models on Bedrock carry Anthropic's marks. | Not confirmed |
| **xAI** (Grok) | No text marking announced. xAI has not signed the Code but remains subject to the Act. | — | No |
| **Stability AI** | — | Invisible watermark plus Content Credentials on Stable Image output. | Not confirmed |
| **Kimi** (Moonshot AI) | No public statement. | — | Not confirmed |
If your institution needs marked text today, pick a Claude model (Fable 5.1 now, Opus 5 from 9 September 2026) or a Gemini model for **Conversation** under Personalization. Assistants can [override the conversation model](/mdx/admin-guide/ai-models#specific-to-each-assistants) individually.
## Checking whether content was generated by AI
Praxis AI does not offer a detector. Detection belongs to the provider that holds the watermark key:
* **Anthropic** runs a **Detection API in private preview** for organisations with a role under EU law — regulators, law enforcement, media, fact-checkers, independent researchers, **educational organisations** and EU civil society groups — and for enterprises that must verify marking for their own compliance with the Act. Request access through Anthropic; see [How Claude marks AI-generated content](https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content).
* **Google** offers the **SynthID Detector** portal to a limited set of partners.
* **Content Credentials** on images and videos (OpenAI, Amazon, Stability AI, Anthropic) can be read with any C2PA-compatible tool, such as the [Content Credentials verifier](https://contentcredentials.org/verify).
A detected mark is a **signal**, not proof. Providers state that a mark shows content was processed by their model, not who wrote it or where the source material came from; short passages and highly factual text may carry too little signal, and heavy rewriting or translation weakens it. Do not treat the presence or absence of a watermark as evidence in an academic-integrity or authorship decision on its own.
## What your institution should do
Provider marking covers the machine-readable side. The obligations that fall on you as the deployer of a Digital Twin are about **labelling** what people see:
1. **Keep the AI disclosure visible** wherever the twin appears — your website, an LMS course, an embedded avatar, a guest link. Do not strip or hide the twin's AI identity.
2. **Label real-person likenesses.** If your twin uses the avatar, image or cloned voice of a real person, make clear that the audio and video are AI-generated. The Act treats undisclosed synthetic likenesses of real people as deepfakes.
3. **Label public-interest publications.** If you publish AI-generated text on matters of public interest (politics, justice, health, fundamental rights) without a human editorial review, say that it was AI-generated.
4. **Know your models.** The [Provider status](#provider-status) table tells you which marks apply to the models your twin uses; review it when you change a model under Personalization.
5. **Ask counsel** about your specific deployment. Praxis AI will answer questions about how the platform behaves at [humans@praxis-ai.com](mailto:humans@praxis-ai.com).
## Related
* [AI Models](/mdx/admin-guide/ai-models) — Model selection, Zero Data Retention, Bring Your Own Model
* [Platform Models & Providers](/mdx/admin-guide/platform-models) — The full provider and model catalog
* [Student Data Privacy](/mdx/introduction/privacy-edu-consent) — Data handling commitments for education deployments
* [WCAG Compliance](/mdx/introduction/wcag-compliance) — Accessibility conformance
## Sources
* Anthropic — [How Claude marks AI-generated content](https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content) and [How Claude's text watermarking works](https://www.anthropic.com/news/claude-text-watermark)
* OpenAI — [EU AI Act: OpenAI resources and customer guidance](https://help.openai.com/en/articles/12141645-eu-ai-act-openai-resources-and-customer-guidance) and [Supporting Europe's work in ensuring a trustworthy AI ecosystem](https://openai.com/index/supporting-eu-trustworthy-ai-ecosystem/)
* Google DeepMind — [SynthID](https://deepmind.google/models/synthid/)
* AWS — [Amazon Nova Canvas](https://docs.aws.amazon.com/ai/responsible-ai/nova-canvas/overview.html) and [Amazon Nova Reel](https://docs.aws.amazon.com/ai/responsible-ai/nova-reel/overview.html) AI Service Cards
* 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), [Code of Practice on Transparency of AI-generated Content](https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content) and the [31 July 2026 signatories announcement](https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content)