Article 50: chatbots, deepfakes and the duty to make AI recognisable

Between the prohibitions (already applicable) and the high-risk regime (expected late 2027) sits Article 50: the transparency layer that touches nearly every organisation deploying AI towards the public. Under the Digital Omnibus agreement of 7 May 2026 this layer becomes applicable on 2 December 2026 — formally, until publication in the Official Journal, the date still reads 2 August 2026 — and the accompanying marking code of practice was published in final form on 10 June 2026.

The four transparency obligations

1. Disclosing interaction with AI: whoever offers an AI system that

communicates directly with people — chatbots, voice assistants — ensures the user knows, or can reasonably know, that they are dealing with AI.

1. Machine-readable marking of synthetic content: providers of

generative systems ensure that output (audio, image, video, text) is detectable as artificially generated or manipulated in a machine-readable format — robust, interoperable and technically feasible.

1. Labelling deepfakes: deployers who generate or manipulate image,

audio or video content that appears authentic disclose that the content is artificial, with a limited exception for evidently artistic and satirical expression.

1. AI text with a public function: text published to inform the public

on matters of public interest is disclosed as AI-generated, unless there is human editorial control and responsibility.

In addition, deployers of emotion recognition and biometric categorisation systems inform the persons concerned — to the extent such systems do not already fall under the Article 5 prohibitions.

The marking code of practice

What "machine-readable marking" looks like technically is being worked out in a code of practice on marking and labelling AI content, steered by the AI Office. The first draft appeared on 17 December 2025; finalisation is expected in June 2026. For providers, this code will be what the GPAI Code is for model builders: not law, but the accepted route to demonstrate compliance — with watermarking, metadata and provenance techniques as the building blocks.

What this means in practice

Article 50 is deliberately light: no conformity assessment, no registration — but visibility. The practical preparation is manageable: make sure every customer-facing chatbot identifies itself as AI; check whether your generative tools support marking and metadata (and switch them on); establish a labelling rule for realistic synthetic image and audio; and document editorial control for AI-assisted publications. Organisations that arrange this now are ready whether the date stays 2 August or becomes 2 December.