EU AI Act Article 50 Checklist: Small Businesses and Creators (US & International)

Who Counts as a Deployer Under Article 50

Article 50 targets providers and deployers of covered AI systems. Solo creators and small businesses that use consumer AI tools (like ChatGPT, Claude, Midjourney, or similar) to publish AI-generated or AI-manipulated content count as deployers of that AI system when they distribute the output online.

The law applies regardless of where a publisher is based. If people in the EU can access a website, YouTube channel, social media posts, or any other publication, that content is published to the EU. US-based creators, UK creators, Australian creators, and publishers from anywhere else fall under this rule if they have online audiences in the EU.

Article 50 applies when content falls into covered categories: text generated by AI and meant to inform the public; images, audio, or video generated or significantly manipulated by AI; or deepfakes (synthetic media that makes someone appear to say or do something they did not). Routine AI use like spell check or grammar tools does not count.

The Disclosure and Labeling Requirements

For deepfakes, disclosure is mandatory. Any synthetic video, audio, or image that makes a real person appear to do something they did not do must be clearly labeled as artificially generated or manipulated. There is no exception.

For AI-generated text meant to inform the public, Article 50 requires disclosure that it was AI-generated, unless a person did substantive review of the content and someone took documented editorial responsibility for it. Editorial responsibility means a real person is named or identified as accountable for the accuracy and integrity of the published text.

For all covered AI outputs, machine-readable marking (metadata) is required so the content can be detected as AI-generated. This is not just a visible label on the page. The metadata must be embedded in or attached to the content so automated detection systems and researchers can identify it. Different formats require different approaches: for images and video, EXIF data or embedded metadata; for web pages, meta tags or schema markup.

The Editorial Review Exception

Article 50 permits an exception for AI-generated text when a person reviews it for accuracy, relevance, and integrity before publication, and someone with editorial authority takes documented responsibility for the published version. Editorial responsibility requires a real person to be named or identified as accountable for the content's accuracy and integrity.

The documented review and responsibility requirement is meaningful. The publisher must have a record showing a real person read the content, made judgment calls about its quality and truthfulness, and is willing to stand behind it. This approach typically applies to publishers, news sites, and organizations with editorial workflows.

For small creators or solo publishers without an editorial team or formal review process, compliance typically requires disclosure of all AI-generated text meant to inform the public.

Deepfakes, Synthetic Media, and Scope

Deepfakes have zero tolerance under Article 50. Synthetic media that shows a real person appearing to say, do, or express something they did not must be labeled as artificially generated or manipulated. No exceptions exist for satire, fiction, or news commentary.

Audio deepfakes are covered. An AI voice that imitates a real person's voice, used in a video or publication that implies the real person said those words, is a deepfake and must be disclosed.

A transition rule provides some flexibility: generative AI systems already on the market before August 2, 2026 have until December 2, 2026 to comply with these rules. If a publisher has been using a particular AI tool before August 2, there is an extra four months. New AI tools or content created after August 2 must comply immediately.

Compliance Steps Before August 2, 2026

Compliance generally involves reviewing recent content. Publishers using AI-generated images, videos, text, or audio should go back through recent posts and identify covered material. Deepfakes require action even before August 2 if any have been published.

Publishers typically must plan their disclosure and metadata approach. For images, this may involve using EXIF data, embedded metadata, or visible labels. For videos, research into embedding metadata or watermarks may be needed. For text, planning where disclosures will appear is necessary. For web pages, adding meta tags or schema markup signals AI generation.

Documented editorial review processes should be prepared. If a person reviews content before publishing, noting who, when, and what they checked helps demonstrate compliance. This matters especially for news, research, or public-facing business content.

Disclosure language should be designed clearly and conspicuously. Examples include: 'This image was generated using Midjourney' or 'This article was written with AI assistance and reviewed by [Name].' Publishers should consult the official EU AI Act Article 50 text and implementation guidance as it becomes available.

Testing machine-readable markup is part of compliance. Publishers who work on websites should add the metadata and verify it is readable. Browser developer tools or metadata validators can confirm the markup is present and correct.

Frequently asked questions