The new rule does not say “slap an AI label on everything.” Marketing teams still need to know which side of the line their content lands on.
On July 20, the European Commission published its final guidelines on Article 50 transparency obligations. The rules apply August 2, 2026. They cover AI systems that interact with people, machine-readable marking by providers, disclosures for deepfakes, and certain AI-generated public-interest text.
This matters outside Europe, too. The Commission’s Article 50 Q&A says providers outside the EU can be covered when system output is used in the EU. A U.S. company with European customers, campaigns, events, or distribution should not assume geography makes the question disappear.
This is a practical marketing triage, not legal advice. If your content or system may be in scope, have counsel confirm the decision.
The hard part is not adding a label. It is knowing who owns the decision before the content ships.
First, separate provider duties from deployer duties
A provider develops an AI system or has one developed and puts it on the EU market under its name. A deployer uses an AI system professionally under its authority. For most marketing departments, the second role is the more likely one.
The distinction prevents a common mess. The AI-tool provider generally owns machine-readable marking of generated output. The organization publishing a qualifying deepfake or public-interest text owns the visible disclosure decision. A metadata marker buried in the file does not replace a clear label when a deployer’s disclosure duty applies.
Run this four-bucket content triage
Bucket 1: Synthetic media that could pass for something real
The Commission defines a deepfake as AI-generated or manipulated image, audio, or video that resembles an existing person, object, place, entity, or event and could falsely appear authentic. Think cloned executive audio, a generated customer scene presented like documentary footage, or a realistic image of a facility that does not exist.
That content needs a clear, perceivable disclosure by first exposure. Do not hide it in a terms page or rely on embedded metadata. Put the notice where a person encounters the asset.
Bucket 2: AI-assisted public-interest text
Published text meant to inform the public on matters such as public health, safety, politics, public services, environmental protection, financial developments, or other subjects of public debate can require a label when it lacks real human review or editorial control.
The useful exception is substantial. Text that receives knowledgeable human review, fact-checking, source review, and approval by someone with authority can fall outside the labeling duty. A spell-check and a quick “looks good” do not count.
Bucket 3: Standard editing and ordinary production help
The Commission says provider marking duties do not apply when an AI system performs an assistive function for standard editing. Routine correction, formatting, cleanup, or production assistance is not automatically the same as generating deceptive synthetic content.
Context still matters. Removing a stray object from a product photo is different from inventing the product. Cleaning room tone is different from cloning a spokesperson. Write down the boundary your team uses.
Bucket 4: AI chatbots, agents, and avatars
Systems designed for direct two-way interaction with people generally need to disclose that the person is interacting with AI from the start, unless that fact is obvious. A background scoring system is different from an AI receptionist answering a prospect’s questions.
Review website chat, lead-qualification agents, automated customer-service avatars, and event kiosks. The disclosure should be clear and accessible before the conversation gets mistaken for a human one.
Build one approval field, not a 40-page policy
Add these questions to the creative brief or publishing ticket:
- Where will this run? Include EU audiences, media, events, and partner distribution.
- What did AI create or materially change? Name the tool and the output.
- Could a reasonable viewer mistake it for a real person, place, object, or event?
- Does the text inform the public on a matter of public interest?
- Who performed substantive human review? Record the reviewer and approval date.
- What disclosure will appear, and where? Save the exact language with the final asset.
Keep the original, prompt or edit note, final export, approval, disclosure language, and published URL together. That extends the practical workflow in our AI content provenance policy and the asset review discipline in our Google Ads AI disclosure guide.
Do not relabel the archive in a panic
The Commission says content generated and already made available before August 2 does not need retroactive labeling, though voluntary disclosure is encouraged. Start with future work and active systems. Then review older high-risk synthetic assets because they can still create trust problems even when retroactive labeling is not required.
The clean operating rule: disclose realistic synthetic media, give public-interest text a real editor, identify AI interactions early, and keep enough documentation to explain the call. Then ask counsel about the edge cases.
Need an AI content review that fits the real workflow?
We can map the tools, approvals, disclosures, and source records your team needs before the next asset goes live.