Marking Synthetic Text Published on Matters of Public Interest

Article 50 addresses not only synthetic video and images but also AI-generated text published on matters of public interest. This includes news articles, policy statements, scientific findings, public health guidance and other content on topics that affect public opinion or policy decisions. A synthetic text requirement differs from synthetic media marking because text generation is scalable, difficult to reverse-engineer and easy to separate from source attribution. The regulation requires machine-readable marking when possible, not just human-visible labels.

The scope is limited to text published on matters of public interest. Private communications, internal documents, customer service responses and entertainment content are outside the scope. A chatbot response to a customer question, even if generated by an AI model, does not require disclosure because the customer knows they are interacting with a chatbot (as addressed in Lesson 2). An employee’s email written with AI assistance does not require disclosure unless the email is published externally on a matter of public concern. Conversely, an AI-generated news article, policy brief or scientific summary published on a website or news feed does require disclosure.

What counts as a matter of public interest

Public interest topics include current events, government policy, public health, environmental issues, elections, regulations, financial markets, scientific research and any matter on which informed public opinion matters. A news article about a disease outbreak is a matter of public interest. A social media post from a government official about a new regulation is a matter of public interest. A financial analyst’s report on a publicly traded company is a matter of public interest. An academic paper published in a journal is not a matter of public interest in this sense (academic communities have different norms and disclosure requirements).

Private opinions, consumer reviews, entertainment criticism and personal blogs are generally not matters of public interest unless they reach mass audiences or influence policy. A restaurant review on a consumer website does not require synthetic text marking. A personal blog post does not require marking. An influencer’s social media post about a consumer product does not require marking unless it is framed as public health guidance or reaches a scale that materially influences consumer behaviour or policy.

Machine-readable marking requirements

Machine-readable marking means metadata or technical signals that automated systems can detect and interpret, without requiring human reading. This includes JSON-LD schema tags, RDF metadata, or proprietary tags used by platforms. An article claiming to be AI-generated should carry metadata stating {“aiGenerated”: true} or similar. The purpose is to allow news aggregators, search engines and AI assistants to identify synthetic content and label it accordingly. Machine-readable marking also prevents manipulation; a human-readable label can be edited or removed, but embedded metadata is harder to tamper with.

Platform responsibilities and organisational disclosure

Platforms that publish content must implement system-level disclosure for AI-generated content. A news website using an AI system to generate article summaries should display “AI-generated summary” prominently and carry machine-readable marking. A social media platform should prompt users who use AI writing tools to mark their posts as synthetic. Publishers have responsibility for disclosure, but platforms should provide tools and nudges to make disclosure easy. If a platform knows a user is using an AI system to generate content, the platform can enforce disclosure as a condition of publication.

Exceptions for simple synthetic text

Short, factual text may not require disclosure. A social media post generated by an AI system that simply states “The stock market closed up 2.3% today” does not require disclosure because the content is purely factual and there is no risk of manipulation. Similarly, system-generated notifications such as “You have a new message” do not require disclosure. The exception applies only to content that is either creative (expressing opinions, arguments or narratives) or where AI generation introduces risk of fabrication or bias.

Evidence of non-disclosure violations

Regulators will look for evidence that text was AI-generated but presented as human-written. This evidence includes metadata in documents, patterns of language consistent with specific AI models, statistical analysis showing unusual consistency in style or argument structure, or admissions by the publisher. Organisations should maintain records showing which articles were AI-generated, when, and under what disclosure protocol. If a regulator discovers that a website published 100 articles on public interest topics using AI generation with no disclosure, the organisation cannot claim “we did not know we had to disclose”.