What applies from August 2, 2026
Article 50 of the AI Act sets transparency obligations for AI systems that interact with people or generate content. Four points matter in everyday life. First: chatbots and voice assistants must tell you that you are talking to an AI, unless that is already obvious. Second: providers of AI systems that generate images, video, audio or text must mark their outputs in a machine-readable format as artificially generated, so software can detect them.
Third, and most relevant for social media: anyone who creates or distributes a deepfake - AI-generated or AI-manipulated image, video or audio that resembles real people, places or events and could falsely appear authentic - must disclose that the content is artificial. Fourth: AI-generated texts that inform the public on matters of public interest must be labeled as AI-generated, unless a human has editorially reviewed them and takes responsibility for publication.
The disclosure must be clear and visible at the latest at the first interaction with or exposure to the content. A note buried in fine print does not satisfy the obligation.
Wasn't the AI Act postponed? Yes - but not this part
In late 2025 the European Commission proposed timeline adjustments to the AI Act as part of the so-called Digital Omnibus, and since then the claim that "the AI Act was postponed" has been circulating. That is only half true. What was postponed are mainly the obligations for high-risk AI systems, for example in hiring or critical infrastructure - they now apply in stages from late 2027 and 2028.
The transparency obligations in Article 50 - exactly the labeling rules for deepfakes and AI content - were explicitly excluded from that postponement. They apply as planned from August 2, 2026. There is a transition period for one sub-aspect only: providers of AI systems that were already on the market before August 2, 2026 get extra time until December 2, 2026 for the technical, machine-readable marking of generated outputs. The disclosure obligation for deepfakes applies from August regardless - without exception. If you assume "everything got delayed anyway", you are wrong at the decisive point.
Who is covered - and who is not
The deepfake disclosure obligation targets so-called deployers: those who use an AI system in a professional or public-facing way and distribute the content - companies, agencies, media outlets, influencers, political actors. If you use AI images in marketing, let AI avatars speak in videos, or post AI-generated scenes that look like real footage, you must disclose it.
Purely private, non-professional use is exempt: sharing an AI image in a family chat is not covered by the regulation. For obviously artistic, satirical or fictional works a softened form applies: the disclosure must exist but must not spoil the work - a piece of satire does not need a disruptive warning banner. Note that the line between private and public-facing is crossed faster than many think: a meme in a small group is not the same as a monetized account with real reach.
Then there is the enforcement reality: legally, the AI Act also reaches providers and distributors outside the EU as soon as their AI systems or content are used here. In practice, though, the obligation is hard to enforce against actors who do not care about European law - they simply will not label, and exactly such content is the rule in disinformation, not the exception.
How to spot a disclosure
Labels will appear in several forms. Most visible are platform labels: TikTok, Instagram and YouTube show notices like "AI-generated" or "Creator labeled as altered" on some content - sometimes because the creator declared it, sometimes because the platform found a machine-readable marker in the file. There are also disclosures inside the content itself, such as an "AI-generated" overlay in a video or a note in the caption.
Invisible but technically most important are machine-readable markers in the file itself: provenance standards like C2PA Content Credentials and watermarks like SynthID, which providers such as Google or OpenAI embed in generated content. You cannot see these signals with the naked eye - you need tools that read them. Our check reads such provenance data where present and shows it in the result. For how these standards work and where their limits are, see our guide on C2PA, SynthID and watermarks.
Why you cannot rely on labels
As sensible as the obligation is, it has structural gaps. The most important one: only content that providers and distributors choose to label gets labeled. Bad actors who want to defraud with deepfakes or spread disinformation will not comply - for them, August 2 changes nothing except the risk of getting caught. Actors outside the EU are legally covered but hard to hold accountable in practice.
Then there are technical gaps. The mass of AI content circulating today was created before the obligation and remains unmarked. Machine-readable markers and metadata frequently get lost in sharing: screenshots, re-uploads and app processing often strip them completely. And providers whose systems were already on the market before the deadline have until December 2026 for the technical marking anyway. In short: a missing label says little about whether content is real.
That is why the second path stays essential: verify content yourself. Watch for the classic visual signs we describe in our guides on AI images and AI videos, and use a technical analysis when in doubt. You can submit a suspicious image, video or link directly to us - we check provenance data, visible and invisible traces, and give you an honest assessment, including the cases where no reliable verdict is possible.
Violations, fines and enforcement
Violating the transparency obligations is not a trivial offense: the AI Act provides for fines of up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher. Enforcement sits with national authorities in each EU country; in Germany, for example, the Federal Network Agency (Bundesnetzagentur) was designated in July 2026 as the central market surveillance and complaints body - citizens can report unlabeled AI content there.
How strictly authorities will enforce the new rules remains to be seen. Realistically, large visible cases will be pursued while the mass of small violations initially will not. Which is one more reason why the labeling obligation is a safety net with holes - and why checking content yourself remains the reliable path. This article provides general information, not legal advice for individual cases.