Disclaimer: I am not a lawyer, and this article does not constitute legal advice. It outlines an engineering and operational framework for implementing Article 50 compliance in real-world workflows.
Article 50 of the EU AI Act (Regulation 2024/1689, amended by the Digital Omnibus, Regulation 2026/1744) has applied since 2 August 2026. Poland’s national AI systems act followed on 11 August 2026 (Dz.U. 2026 poz. 1003), and the part that matters operationally, inspections and fines, starts 28 October 2026, run by the new supervisory body KRiBSI with appeals going to the SOKiK court. Outside that track, UOKiK, media regulators, and plain civil claims over likeness or copyright were already live before any of this.
Most teams I talk to are stuck between two wrong defaults: label every AI-touched asset, or assume a disclaimer in the footer covers everything. Neither survives contact with an inspection or a court. What decides is narrower and more mechanical than either instinct: your role (provider vs. deployer), the content type, the context it’s published in, and whether a human actually controlled the output. The name of the tool you used answers none of it.
The transitional period is not a grace period
There’s a specific misreading going around: “we have until December, so no rush.” The transitional window runs until 2 December 2026, but it covers one thing: technical marking under Art. 50(2), for providers whose systems were already on the market before 2 August 2026. It says nothing about deepfake labeling or public-interest text disclosure. If you’re a deployer publishing a synthetic video or an AI-drafted article today, the obligation is live now, not in December.
There’s no retroactive duty either. Content you published before 2 August doesn’t need a label bolted on after the fact. But republishing counts as a new act. If you pull an old AI-generated clip back onto a new channel, or repost an old article on a new platform, that’s a fresh publication and needs a fresh assessment against current rules, not a pass because the original upload predates the law.
Roles matter more than people expect. Prompting a model, running personas, or wiring an API into your product usually keeps you a deployer. You become a provider when you put a system on the market under your own brand; at that point you inherit the full provider obligation set, including the Art. 50(2) marking duty. An employee working inside your organization is not a separate deployer from the company; the obligation sits with the org.
Three tests decide deepfake labeling, not vibes
For synthetic image, audio, or video, the AI Act’s deepfake definition is cumulative. All three conditions have to hold before disclosure is required:
- The content is AI-generated or AI-manipulated.
- It resembles an existing person, object, place, entity, or event.
- In context, it could falsely appear authentic to an average recipient in the actual channel where it’s published.
Fail any one test and it’s not a deepfake under the Act, even if it’s obviously synthetic to a trained eye; the standard is the average viewer on that platform, not an expert reviewer. Pass all three, and you owe visible or audible disclosure at first exposure. Consent from the person whose likeness you used does not remove this: consent and transparency are separate obligations. Art, satire, and content that’s evidently fictional can use a lighter disclosure form that doesn’t wreck the work, but that’s a contextual judgment, not an automatic exemption you get by calling something “creative.” If you can’t state in one sentence why an average viewer in that channel would read it as fiction, treat it as a regular deepfake candidate and disclose.
Concrete calls that come up in practice: a cloned voice of a real host reading copy they never recorded gets a label. A CEO avatar delivering lines the CEO never said gets a label. Generated fire, a fake protest, a fabricated break-in styled like news footage: label it, full stop.
A real product shot against an AI-generated background, product untouched, needs no label as long as no product feature changed. The moment AI adds a feature or effect the product doesn’t actually have, you’ve crossed into labeling territory and picked up a misleading-advertising risk on top. For dubbing: straight translation of a real recording needs no label; a cloned voice standing in for a specific real person triggers the deepfake test again. If you can’t defend the sentence “this is only a technical correction,” treat the change as significant and run the full three-test assessment.
Assisted, manipulated, generated, and the control question
The taxonomy that actually holds up under review has three tiers, and the boundary between them is one question: did AI change the meaning, the evidence, or the apparent authenticity of what’s being shown?
Assisted: AI helped with research, typo fixes, formatting, cropping, color correction. Nothing about what the content claims or shows has changed. No visible label needed in most cases.
Manipulated: AI changed an argument, a voice, a scene, or how a product looks or works. This tier needs an actual assessment, not a reflex. Ask what the change did to meaning, evidence, or apparent authenticity, and decide from there.
Generated: AI created new text, image, audio, video, a person, or an event that didn’t exist as source material. Context still decides the label, not the fact of generation alone. A generated hero image for a blog post is different from a generated “photo” of a real event.
Public-interest content (health, consumer safety, environment, regulation, cybersecurity, major economic events) published to inform the public must be labeled if AI wrote it, unless the editorial-review exception applies. That exception has a real bar: a named human opened the sources, verified the claims, checked dates, numbers, and names, could have rejected the piece, and takes responsibility for it. “I skimmed it and it looked fine” doesn’t clear that bar. An AI-written piece on a security vulnerability with no real review needs a label; the same piece after documented editorial review can rely on the exception. An ordinary ad doesn’t become public-interest content just because AI drafted the copy.
Building the pipeline that survives an inspection
Disclosure mechanics under Art. 50(5) are specific: clear, distinguishable from the rest of the content, present no later than first interaction or first exposure, and meeting accessibility requirements. Machine-readable marking (metadata, watermarking) is the provider’s job. Human-perceivable disclosure for deepfakes and public-interest text is the deployer’s job, and metadata alone doesn’t satisfy it if a recipient would need a special tool to read it. The EU Commission publishes free disclosure icons; using them is optional, but the underlying disclosure duty holds regardless.
Labels don’t survive channels by default. Platform re-encoding, cropping, and reposting strip metadata and visual context routinely. After you publish, open the material the way an actual user would, on mobile and desktop, and screenshot it. When content moves to a new channel, re-run the assessment; don’t assume the label traveled with the file. This is the same discipline covered in blocking quality gates in content pipelines — a check that runs once at generation time and never again isn’t a check, it’s a false sense of coverage.
Agency and freelancer handoffs are where most of this quietly breaks. Contracts should require disclosure of which systems were used, proof of rights and consents, a named person who decides, places, and verifies labels, no metadata stripping without notice, delivery of evidence files alongside the deliverable, and your right to hold publication if something’s unclear. If you’re an agency, build these AI disclosure terms into your standard contract now rather than negotiating them mid-crisis.
Keep a publication-decision register: ID, material or URL, system and version, AI-use tier (assisted/manipulated/generated), deepfake test result, decision, reviewer or owner, and a post-publication screenshot. Keep it proportionate; don’t archive confidential prompts to prove compliance, just the decision trail. This register is exactly what KRiBSI will ask for during an inspection, and it’s what typically justifies a reduced fine when something slips through despite a documented process. Every AI-touched publication needs a named owner; “marketing posted it” isn’t one.
For practical case breakdowns across video, audio, and marketing assets (in Polish), Tomasz Turba breaks down real-world labeling scenarios in his Sekurak training: „Jak oznaczać treści AI, by nie dostać kary? AI Act na przykładach” on SekurakTV.
For teams starting from zero, a realistic rollout fits in a week: inventory your systems, channels, and agencies (day 1), map provider vs. deployer roles per system (day 2), apply the assisted/manipulated/generated taxonomy to your current content types (day 3), write label copy and decide placement per channel (day 4), wire in review and the evidence register (day 5), update contracts with agencies and freelancers (day 6), then dry-run a real publication end to end (day 7). None of this requires new tooling, just someone deciding in writing who owns the call.