EU AI Act Transparency Rules Now Apply to Chatbots and AI Content

Chatbot and synthetic-content duties are live

EU AI Act transparency rules started applying on 2 August 2026. The European Commission's guidance covers interactive AI and generated or manipulated content, including duties that affect chatbots, image and audio generators, deepfakes, and some AI-written public-interest material.

An AI system designed to interact directly with people must inform them that they are dealing with AI, unless that is already obvious to a reasonably informed and observant person in the circumstances. This is a context test, not a rule requiring the same warning on every interface.

Providers of systems that generate synthetic text, audio, images, or video must also make outputs detectable in a machine-readable format, as far as technically feasible. The law calls for methods that are effective, interoperable, reliable, and robust. It does not prescribe one universal watermark.

Visible disclosures depend on the use

Human-facing disclosure duties differ from the provider's technical marking duty. A deployer must disclose deepfake content. AI-generated or manipulated text published to inform the public on matters of public interest also needs disclosure, unless it has undergone human review or editorial control and a person or organisation holds editorial responsibility.

People exposed to emotion-recognition or biometric-categorisation systems must be informed about their operation. Exceptions and special rules apply, including for some authorised criminal-law uses.

These distinctions matter because “AI-generated content must be labelled” is too broad. The answer depends on who provides or deploys the system, what the content depicts, how it is published, whether meaningful editorial review occurred, and whether an exception applies.

Claude shows how a provider is applying the rule

Anthropic has now described how supported Claude models will mark generated content. Models launched in the EU on or after 2 August 2026 will embed an imperceptible watermark in generated text. Supported files such as SVG, PNG, and JPEG will carry signed provenance metadata using the C2PA standard. Anthropic says the marking applies worldwide across its products and cloud partners where the relevant technique is supported.

The mark is evidence that content may have been processed by Claude, not proof that Claude wrote the underlying ideas or that the material is accurate. Proofreading, translation, formatting, or file conversion can produce a marked output based on human work. The reverse is also possible: heavy editing, short passages, translation, removed metadata, older models, or unsupported platforms can leave no detectable mark. Anthropic has not yet published its promised third-party detection mechanism, so the current announcement describes implementation and limits rather than a complete public verification tool.

High-risk deadlines moved, transparency did not

Recent EU amendments delayed many obligations for high-risk AI systems. Rules for systems listed in Annex III, including certain uses in employment, education, credit, law enforcement, and essential services, are now scheduled for 2 December 2027. Rules for high-risk AI embedded in regulated products are scheduled for 2 August 2028.

That delay did not remove the transparency rules that began on 2 August 2026. A narrower transition applies to the machine-readable marking duty: providers of relevant systems placed on the market before that date have until 2 December 2026 to comply.

The European Commission's enforcement powers for general-purpose AI model obligations also began on 2 August. Those obligations have applied to newer model providers since August 2025, while qualifying models placed on the market earlier have a separate transition until August 2027.

Local processing is not a blanket exemption

Running a model on a phone or computer changes the data path, but it does not by itself settle the AI Act analysis. A locally generated answer can still come from an interactive AI system or be published as synthetic content. Conversely, a clearly presented private chatbot may satisfy the interaction notice test because its AI nature is obvious, although other duties can still apply.

Developers should record which legal role they occupy, where the product is offered, what each feature generates, and how disclosure or machine-readable marking is implemented. Publishers should keep evidence of human review and editorial responsibility when relying on that provision.

For users, the immediate change is more modest. Expect clearer notices when interacting with AI and more provenance signals attached to generated media. Those signals describe how content was created. They do not prove that an output is accurate, safe, privately processed, or free from manipulation.