New European Union regulations now mandate that AI-generated images, videos, and text must bear machine-readable markers, marking a significant shift towards transparency in digital content and posing new challenges for social platforms and content authenticity.
Machine-generated images, video and text are now meant to be recognisable as such in the European Union, while users should also be told when they are dealing with an AI system rather than a person. The new transparency rules took effect this month and extend the wider framework of the EU AI Act, signalling a shift from voluntary labelling towards enforceable obligations.
The change matters because generative AI has already blurred a long-held assumption: that what appears on screen is trustworthy simply because it looks real. According to the European Commission, Article 50 of the AI Act introduces transparency duties so people can identify when they are interacting with AI or viewing content generated or manipulated by it. In practice, that means the origin of an image, clip or passage of text should be machine-readable and, where possible, verifiable.
Major AI companies have begun adapting. Anthropic said its Claude models released from 2 August now include invisible watermarks in generated text, while the company also plans provenance markers for images and tools for detection. OpenAI says it already uses provenance signals such as C2PA Content Credentials and SynthID watermarks for images produced with its tools, drawing on an industry standard designed to carry information about how digital content was made.
The technical details matter. SynthID, developed by Google DeepMind, embeds an imperceptible watermark directly into generated media, while C2PA acts more like a digital passport, storing metadata that platforms can inspect. The key limitation is that the presence of a watermark does not help unless social networks and other services actually check for it. Researchers and policy advisers say the real test will be whether platforms can surface those signals to end users in a clear way.
That is where the rules could have broader consequences than simple labelling. Walter Pasquarelli, a synthetic-content researcher at the University of Cambridge and adviser to the OECD, has argued that AI companies will need to work with the platforms where their output is shared. If they do, a post created in Claude or another system could carry an embedded signal that a social network recognises and labels before publication. That may affect not only deception and misinformation, but also the wider market for mass-produced, low-quality content increasingly described as “AI slop”. LinkedIn has already introduced a way for users to flag posts that appear to fit that description, suggesting the debate is moving beyond fraud and into everyday content quality.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
