EU AI Act forces social media to label AI content from 2 August, but most marketers still don't know what counts

Article 50 of the EU AI Act applies from 2 August 2026, and it forces companies to label certain AI-generated content, most notably deepfakes and AI chatbots. Breaches can draw fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher, and the rules apply to any content directed at EU users, regardless of where the brand is based.

Three weeks in, one pattern is clear: confusion often arises when teams mix up legal requirements, platform rules, and internal policies.

A social media manager using Midjourney for an Instagram carousel, a brand publishing AI-voiced TikTok ads, and an agency drafting LinkedIn posts with ChatGPT all face the same question now, and most still don't have a definitive answer: does this legally require a label, or does turning on the platform's AI toggle cover it?

The short version is that the two layers sit on top of each other, and neither replaces the other.

What Article 50 actually requires (and what it doesn't)

Article 50 requires providers and deployers of AI systems to be transparent about the use of AI in four key areas: direct interaction with individuals, AI-generated content, emotion recognition and biometric categorisation, and deepfakes and AI-generated text on public-interest matters.

For social media marketers, two of those matter most.

Deepfakes. The new EU AI labeling duties reach deepfakes, synthetic photographs, cloned voices, generated video, and AI-written text published to inform the public on matters of public interest. Deployers who use AI to generate or manipulate image, audio, or video content that qualifies as a deepfake must disclose that the content is artificial.

The definition is broader than most teams expect. It covers not only fake politicians or cloned voices, but also realistic-looking AI product images, fabricated press photos or synthetic stock images that appear authentic.

If it looks real enough to fool someone, it needs a visible label. Could this content actually exist in the real world? A photo-realistic portrait of a fictional person is a deepfake. An elephant driving a sports car or a sphinx flying over the Eiffel Tower is not. These scenarios obviously go against the laws of physics and so you don't need to label them.

Public-interest text. Deployers of generative AI systems must clearly label AI-generated or manipulated text published with the purpose of informing the public on matters of public interest. Financial, political, scientific, or cultural developments that may be relevant subject of public debate fall into this category.

Most everyday social media captions do not. The Article 50 text-labeling duty applies to text published to inform the public on matters of public interest, such as politics, public health, or safety, not to promotional captions or product descriptions.

The editorial exemption (and why it matters more than most realise)

Here is where the noise falls away: if a human with editorial responsibility reviewed it before publishing, the text-labeling duty never applies, and the review must check the substance, not just running spell-check.

Human review refers to the deliberate examination of the substance of the content by one or more natural persons possessing relevant knowledge and professional judgement pertaining to the subject matter under scrutiny. A specific person or company assumes full editorial responsibility for the publication.

In practice: AI-drafted text plus real human review equals no label required. Ever.

The catch: one thing the editorial exemption does not do is extend to images, audio, or video. A deepfake needs a visible disclosure no matter how many people reviewed and approved it before publishing.

The editorial-responsibility exemption is generous, and most established outlets already have a named editor in the loop. EU AI labeling risk sits instead with high-volume, low-touch content operations: aggregators, SEO farms, and automated local news feeds where no human meaningfully reviews the output before publication.

Platform labels versus legal compliance

Platforms will automatically label your work as AI-generated. For example, LinkedIn applies a CR label to AI images to ensure transparency and trust on the platform. Other social media platforms like Instagram and TikTok are doing the same and encourage their creators to use the relevant labels.

The problem: platform labels help, but detection is inconsistent and metadata can be lost when content moves between systems. Article 50 places the disclosure duty on you as the deployer, so you shouldn't rely only on a platform's automatic tag for content that legally requires a visible or audible label.

There is no single AI ad label rule in 2026. Meta, TikTok, and Google each define what counts as AI, who applies the label, and what happens if you skip it differently.

This divergence creates a compliance challenge for any advertiser operating across multiple platforms. An AI-generated ad creative that is fully compliant on one platform may violate the rules on another.

What brands must do: check whether the content involves realistic AI media, a deepfake, or AI-generated text on a matter of public interest. Confirm which labeling feature or disclosure each social media channel requires. Reviewing a post for Instagram does not automatically cover TikTok, YouTube, or LinkedIn.

What needs a label on social media, in practice

Run through this checklist for every piece of AI-assisted content before it goes live:

Does it look photorealistic and could a viewer think it's real? If yes, and AI generated or substantially altered it, label it as a deepfake. This includes synthetic influencers, AI-generated product shots that look like photographs, voice clones, and generated video of real locations or events that never happened.

Is it text published to inform the public on a political, health, financial, scientific, or cultural matter? If yes, and a named human did not review and take editorial responsibility for it, label it.

Is it a chatbot or voice assistant? If people interact with an AI chatbot or voice bot on your channels, they must be told they're dealing with an AI at the start of the interaction, unless it's already obvious. A short line such as "Hi, I'm an AI assistant" is enough.

Is it a promotional caption, a product description, or everyday marketing copy that a human reviewed? No label required under Article 50.

Did AI help with editing, translation, or brainstorming, but a human created the substance? Brainstorming support, translation, or an editor-reviewed caption does not automatically need a label.

How to label (when you have to)

The European Commission adopted guidelines on these obligations on 20 July 2026, and the EU has developed a set of icons that creators, publishers and other deployers of generative AI systems may use to label their AI-generated content. These icons are freely available and an integral part of Section 2 of the Code of Practice on marking and labelling of AI-generated content.

The above information must reach the user in a clear, distinct manner, and at the latest during the first interaction or first contact with the generated content. It is also required that these labels meet accessibility standards for individuals with disabilities.

In practice, "AI-generated image" or "Made with AI" in a visible, accessible format covers most use cases. Performance improved across all measures when the basic icon was accompanied by a text label (e.g. modified).

Where the confusion will stay for a while

Three gaps remain undefined, and teams will have to make judgment calls until enforcement and case law settle them:

What counts as "substantially altered"? AI editing is fine under Article 50 except when the AI "substantially" alters the input data. Well, what constitutes substantially? That's not defined by Article 50, either. Routine edits like colour and lighting correction, noise reduction, cropping, compression, accessibility tweaks generally don't require a label.

How rigorous must human review be to qualify for the editorial exemption? The law does not say what constitutes a human review. Is it enough to give it a passing glance? The safer interpretation: a meaningful, substantive review by someone with subject-matter knowledge who can point to what they checked and why they approved it.

What exactly is a "matter of public interest"? Are review and recipe sites public interest sites or does that only apply to political sites? Article 50 doesn't explain any of that. The Commission's guidance suggests political, health, financial, scientific, and cultural topics, but the edges remain blurry.

When in doubt, label. A small visible note like "AI-generated image" is cheap insurance and, increasingly, something audiences respect.

What social media teams should do now

The EU transparency obligations apply from August 2, 2026, but effective workflows do not appear overnight. Social media teams can use the time remaining to turn one-off decisions into a clear, repeatable process.

Map every AI tool the team uses. List every generative system your organisation provides or deploys into the EU, note the modality, and record whether it produces authentic-looking output. You cannot label what you have not counted, and most inventories surface tools nobody in legal knew existed.

Define three tiers. Green: minor AI help, no label needed. Yellow: realistic AI generation, add the platform toggle plus a short caption line. Red: deepfakes, synthetic humans, AI applied to public-interest content (get sign-off before publishing).

Document editorial responsibility for text. If you're relying on the editorial exemption, name who reviewed what, and keep a record. It is mandatory to implement proper human review and editorial control procedures, explicitly identifying a specific natural or legal person who carries full editorial responsibility for the publication.

Turn on platform AI toggles by default, but don't stop there. The platform label is helpful but not sufficient for legal compliance where Article 50 applies. Add your own visible disclosure for in-scope content.

Review quarterly. The rules are still changing, and platform policies update faster than legislation. What's compliant today may not be in three months.