This article is auto-synced from its in-app version in Tai.

The EU AI Act is the European Union's law on artificial intelligence. This page covers the parts that concern TAI, and how TAI meets them.

This page explains how TAI works. It is not legal advice.

When you'd use this

Read this when your legal or compliance team asks how TAI handles the AI Act, or before you publish a report that TAI produced.

The AI Act in short

The law. Regulation (EU) 2024/1689. Its transparency rules are in Article 50.
When. Article 50 applies from 2 August 2026. For AI systems already on the market before that date, the machine-readable marking duty applies from 2 December 2026.
Who. Nezasa provides TAI, so Nezasa is the provider. Your company uses TAI under its own authority, so it is a deployer.
What Article 50 asks:
- Providers tell people they are interacting with an AI system, unless that is obvious (Article 50(1)).
- Providers mark AI-generated content in a machine-readable format, so it can be detected as AI-generated (Article 50(2)).
- Deployers who publish AI-generated text to inform the public on matters of public interest disclose that it is AI-generated (Article 50(4)). This does not apply when the text has gone through human review or editorial control, and a person or company holds editorial responsibility for the publication.

How TAI addresses it

Chats

Every chat says You're chatting with AI. — under the empty-chat message, and as a caption under the message box. It appears in all 14 TAI languages, and in chats embedded in other apps through the Chat UI SDK once the app runs the current SDK version. Developers can reword it, but not remove it.

An open chat with the TAI Assistant, with the notice "You're chatting with AI." under the message box

Reports

The report viewer, the reports library and the standalone view label every report AI-generated, next to the agent's name.
Export PDF and Export HTML end with a footer saying Generated by AI., in your language. The PDF footer also numbers the pages.
The same footer is on PDFs you export from a chat's HTML preview, and on the PDF attached to scheduled report emails.

Machine-readable marking

HTML files TAI generates carry two meta tags: tai-ai-generated (set to true) and tai-generator (TAI and the agent's name).
Exported PDFs carry an AI-generated flag in their document metadata.

None of these labels or markers can be switched off.

When you publish a report

Keep the AI-generated label, the footer and the metadata in place.
If you publish a report as-is on a matter of public interest, Article 50(4) asks you to disclose that it is AI-generated. The footer does that.
If someone on your side reviews the text and your company takes editorial responsibility for it, the labelling duty may not apply. Your legal team decides which case you are in.

Why TAI doesn't watermark text itself

A model's text watermark is invisible. The model weaves it into the text as it writes, so only the model provider can add it — not an app like TAI that uses the model.

TAI uses AI models from model providers such as Anthropic. Anthropic watermarks text from its supported Claude models at the model level. Coverage depends on the model and on where it runs: on cloud platforms such as Amazon Bedrock, which TAI uses, only Anthropic's newer models carry it. Claude models launched in the EU from 2 August 2026 support it at launch, and Anthropic is adding it to older models during a transition period. Heavy editing, paraphrasing or translation can remove a watermark.

Two EU codes — not to be confused

General-Purpose AI Code of Practice (published July 2025). A voluntary code for providers of general-purpose AI models. Signing it lets them show they meet the AI Act's model obligations on transparency, copyright, and safety and security. Signatories include Anthropic, Amazon, Google, Microsoft, Mistral AI and OpenAI.
Code of Practice on Transparency of AI-generated Content (published June 2026). A voluntary code for Article 50: marking by providers, labelling by deployers. It expects marking in layers — signed metadata plus an invisible watermark. Free-form text cannot carry metadata, so for text a watermark alone is enough. Marking may be done upstream, by the model provider; a provider building on that model may rely on it, but stays responsible for its own output. Anthropic has signed this code.

TAI's share is the visible notices and labels, the meta tags and the PDF metadata. The text watermark comes from the model provider.

Tips and limits

TAI does not offer a watermark detector. Anthropic gives detection access to eligible organisations, as EU law requires.
The meta tags and PDF metadata live in the files TAI produces. Text you copy out of a report or chat leaves them behind; only a model watermark, where present, travels with it.

EU AI Act