EU AI Content Labels: A Decision Guide for Communications Teams

EU AI Content Labels: A Decision Guide for Communications Teams

What changed on 2 August 2026?

Article 50 of the EU AI Act now requires specific forms of AI interaction and synthetic content to be disclosed or technically marked. The rules have applied since 2 August 2026, so communications teams need a publishing decision that distinguishes ordinary assistance from deceptive synthetic media, public-interest text and direct AI conversations.

The law separates providers from deployers. A provider develops an AI system, or has one developed, and places it on the EU market under its own name. A deployer is the person or organisation using an AI system under its authority for professional purposes. For most brand, agency and corporate communications teams, the deployer questions are the immediate ones: what was created, what might an audience believe, who controlled the tool and who accepted responsibility for publication?

Geography does not end the assessment. the European Commission’s Article 50 questions and answers explain that a provider outside the EU may still fall within scope when its system output is used in the Union. A campaign produced in Istanbul but distributed to audiences in Paris, Berlin or Amsterdam therefore needs an EU-facing review rather than a simple “created outside Europe” exemption.

Which communications assets require a clear label?

Realistic deepfakes, AI-generated or manipulated public-interest text without qualifying human oversight, and direct AI interactions are the communications assets most likely to require a clear disclosure. The obligation is tied to content, context and audience expectations rather than to the file extension or the name of the software used.

Under Article 50 of Regulation (EU) 2024/1689, a deepfake is AI-generated or manipulated image, audio or video that resembles an existing person, object, place, entity or event and could falsely appear authentic or truthful. A synthetic voice note in which a real chief executive appears to announce an acquisition is a clear risk. A visibly fictional animation that no reasonable viewer would mistake for a recording of a real event calls for a different, proportionate disclosure.

A usable triage starts with four content classes:

• Interactive AI: A chatbot, agent or avatar designed for a genuine two-way exchange should identify itself at the start unless the interaction is already obvious to a reasonably informed person.
• Deepfake media: A realistic imitation of an existing or plausibly existing person, place, object, organisation or event needs a visible or audible disclosure at first exposure.
• Public-interest text: Informative text about issues such as public health, consumer safety, environmental protection, finance, science or culture needs a label when it lacks qualifying human review and editorial responsibility.
• Assistive editing: Routine correction that does not substantially alter the input or its meaning may fall outside the provider marking duty, but the team should still record what the tool changed.

Teams should not treat “creative campaign” as a universal safe harbour. The Act allows artistic, satirical, fictional and comparable works to disclose synthetic manipulation in a way that does not spoil the experience. It does not erase disclosure altogether. A label can be proportionate and still be unambiguous.

When does human review remove the text-labelling duty?

Human review can remove the specific labelling duty for public-interest text only when knowledgeable people examine the substance and a person or organisation holds final editorial responsibility. A spelling pass, tone adjustment or automatic fact-check badge is not enough.

The Commission describes human review as a deliberate examination by people with relevant knowledge and professional judgement. Editorial control also requires practical authority to approve, change or reject the substance, including checking facts and the trustworthiness of sources. This is a governance test, not a credit line placed beneath the finished article.

A strong record should answer five questions. Who reviewed the claims? Which original sources were opened? What significant changes were made? Who could stop publication? Which natural or legal person accepted ultimate responsibility? If the answers exist only in chat messages scattered across several tools, the organisation may struggle to demonstrate that the review was meaningful.

The text exception does not automatically settle the visual assets around it. A thoroughly reviewed article can still include a synthetic interview clip that resembles a real person. That clip may need a deepfake disclosure even when the article copy is published under full human editorial control. A channel-by-channel asset register prevents one approval from being stretched across unrelated risks.

Where should an AI disclosure appear?

An AI disclosure should be clear, distinguishable, accessible and present no later than the audience’s first interaction or exposure. Hidden metadata can support provenance, but it does not replace a human-readable or audible label for deployer obligations.

Placement follows the point of encounter. A video disclosure belongs in the opening frame or another immediately perceivable position, not only in a description that disappears when the video is embedded. An audio disclosure should be heard before the synthetic voice can be mistaken for a real speaker. A chatbot notice should appear in the first conversational state, before the visitor supplies information or acts on an answer.

Machine-readable marking is a different layer. Providers of generative AI systems must make relevant outputs detectable as artificially generated or manipulated through effective, interoperable, robust and reliable technical measures where feasible. Communications teams should preserve those signals when resizing, transcoding or distributing an asset, but they must also add the visible disclosure required for the audience.

Before release, test the label in the actual delivery conditions:

• Check portrait and landscape crops on a small mobile screen.
• Play video without sound and audio without a companion page.
• Embed the asset on a third-party page and confirm the disclosure survives.
• Review colour contrast, reading order and screen-reader access.
• Verify that the label accurately says generated, manipulated or AI-operated.

Vague language creates avoidable doubt. “Digitally enhanced” does not tell an audience whether an ordinary photograph was colour-corrected or whether a person and event were invented. Use direct language that matches the intervention: “This image was generated with AI,” “This video was altered with AI,” or “You are interacting with an AI assistant.” Localise the wording for each market instead of translating a legal sentence word for word.

How should the publishing workflow change?

The publishing workflow should classify the asset, document the source, assign an accountable reviewer and test the disclosure before a channel owner can release it. Starting with a list of approved AI tools is useful for security, but it does not answer the audience-facing questions in Article 50.

Begin with a focused inventory covering recent and planned communications. Include press materials, social assets, executive videos, customer-service agents, campaign pages and reusable agency stock. Record the original input, the system, the material change, target market, channel and first-exposure point. Keep this legal review distinct from the broader work of building brand authority in the AI era.

Then give every asset one owner. The Commission explains that a legal person can remain the deployer when employees, contractors or freelancers operate a system on its behalf and under its control. Contracts should therefore identify who classifies content, who preserves provider markings, who writes the visible label, who tests channel placement and who can halt publication.

A compact release record can use these fields:

• Source asset, creation date and rights holder.
• AI system and the substantive transformation performed.
• Audience, market, channel and expected context.
• Disclosure decision, wording and first-exposure placement.
• Human reviewer, legal check and final editorial owner.

This register also strengthens incident response. If a fake executive statement circulates, the team can separate verified originals from unauthorised versions, brief journalists with consistent evidence and issue a correction without first reconstructing its production history. A connected source record supports evidence-led earned-media authority by making claims traceable to accountable sources.

Evidence: FL PR & Communications presents its own crisis practice as a sequence of planning, intervention and repair. Its official Instagram posts stating that every market asks a different trust question and that traffic changes source rather than simply disappearing are first-party signals behind this guide’s market-specific disclosure and provenance workflow. They are public descriptions of the agency’s approach, not earned-media results.

The Code of Practice on Transparency of AI-generated Content offers a voluntary route for demonstrating compliance with the marking and labelling duties. The legal obligations are not voluntary. Organisations that do not use the code still need measures they can defend as adequate, and the Commission notes that alternative approaches may attract more requests for information.

What should a Türkiye-based team do for EU campaigns?

A Türkiye-based team should build one controlled source package and separate release instructions for every EU language, market and channel. Central approval in Istanbul does not guarantee that a German customer, French journalist and English-speaking visitor encounter the same disclosure. The discipline also explains why international press release distribution alone is not enough: each market still needs accountable local presentation.

Use a market sheet that travels with the master asset. It should name the local disclosure, the placement, the responsible publisher and any sector review. A medical spokesperson’s synthetic voice, an investor-facing executive clip and a fictional fashion film deserve different escalation paths because the foreseeable harm and audience expectation are different.

Local language matters because transparency is an understanding test. An unexplained English “AI-generated” badge in Turkish copy may not tell every reader whether the entire work was created synthetically or only one element was changed. Conversely, a literal translation of a Turkish compliance sentence can sound evasive in English. Native editors should produce short, direct labels that preserve the legal meaning.

Do not outsource responsibility to a platform badge. Platform labels can vary by country, disappear in an embed or be lost when another user downloads and reposts the file. Preserve three distinct layers: provider provenance signals inside the file where available, a clear audience-facing disclosure, and an internal editorial record tying the released version to its source.

The final test is straightforward: if a reasonable audience could mistake synthetic content for an authentic person, event or statement, place transparency before the creative reveal. If the team relies on human editorial control for public-interest text, preserve evidence of substantive review and final responsibility. Trust is easier to protect when the disclosure decision is made before distribution, not after a challenge appears.

Frequently Asked Questions

These answers cover the disclosure, review and accountability decisions communications teams are most likely to face.