What does the EU AI Act require for AI-generated video?
On 2 August 2026, the transparency rules in Article 50 of the EU AI Act enter into force (EU AI Act, Article 50, 2026). They land just as AI video becomes a routine tool for communications teams, and they change AI video from a free-for-all into a governed activity. The deadline is roughly six weeks away, so the time to map your exposure is now, not in late July.
Two obligations matter most for comms teams. First, providers of AI systems that generate synthetic audio, image, video, or text must mark the outputs in a machine-readable format so they are detectable as artificially generated (EU AI Act, Article 50(2), 2026). Second, deployers who publish a deep fake, meaning AI-generated or manipulated video of a realistic person, must disclose that the content is artificially generated (EU AI Act, Article 50(4), 2026). The disclosure has to be clear and given at the latest at the first moment someone sees the content.
There is nuance worth knowing. The marking duty does not apply where AI only assists standard editing or does not substantially alter the material (EU AI Act, Article 50(2), 2026). And AI-generated text that informs the public is exempt from disclosure when a human reviewed it and a named person holds editorial responsibility (EU AI Act, Article 50(4), 2026). A provisional Digital Omnibus agreement would also give providers already on the market before August a transitional period until 2 December 2026 (Global Policy Watch, 2026).
Why is this landing right now?
Because adoption and regulation arrived in the same quarter. More than half of internal communications practitioners, 52 percent, now use AI at least a few times a week, and 53 percent report a very high or significant impact on their efficiency (Simpplr State of Internal Communications, 2026). AI is no longer an experiment in the comms function; it is in the weekly workflow.
The regulation is not arriving in a vacuum either. The EU is finalising a voluntary Code of Practice on transparency, expected in June 2026, and a provisional Digital Omnibus agreement is already negotiating transitional timelines into late 2026 (European Commission, 2026). The rules are being turned into working practice right now, which is why comms teams cannot file Article 50 under "next year".
So the pressure is real on both sides. Teams want to publish faster with AI video, and the law now asks them to prove how that video was made. Treating these as two separate problems, adopt now and comply later, is how teams stall on both.
How can comms teams stay compliant without slowing down?
The teams that keep their speed treat compliance as a property of the tools they choose, not a manual checklist they run after the fact. Governance built into the production pipeline clears the deadline without adding review rounds. Three practical moves cover most of the ground.
Separate real footage from synthetic presenters. A real employee who records themselves and consents is not a deep fake, so a large share of internal comms video sits outside the heaviest disclosure category. Ella (guided employee recording) is built for exactly this: real people, on brand, with consent baked into the workflow. Where you do use synthetic presenters, like Theo (text-to-video), label and disclose them, and back that with a clear avatar and likeness policy. Digital replicas and synthetic likenesses are now their own governance category, so a written likeness policy is no longer optional.
Make brand and control non-negotiable at the point of creation. Article 50 is, at heart, about auditability and control. That is the same instinct behind Brand Templates: every output from Milo, Ella, and Theo is brand-compliant by construction, so content cannot leave the brand box. A pipeline where outputs are controlled and traceable by default is far easier to bring into line with a transparency rule than a pile of ad hoc clips from a dozen free tools.
Inventory where AI touches your video, then write it down. List every place AI generates or alters a clip, decide which outputs need marking or disclosure, and assign a named owner for editorial responsibility. The voluntary Code of Practice expected in June 2026 will give you a template to demonstrate compliance (European Commission, 2026). Pair that with vendors who can show their certifications. cofenster is ISO 27001:2022 certified, GDPR compliant, and EU AI Act compliant, which is the kind of evidence procurement and legal will ask for.
What does this look like in the DACH market?
In Germany, Austria, and Switzerland the bar is higher, and that is an advantage for teams that prepare. The EU AI Act applies as the KI-Verordnung, DSGVO already governs how you process employee data, and deploying AI that records or replicates employees often triggers Betriebsrat co-determination. The avatar consent question is not abstract here; it is a works-council conversation.
This is why governance-led DACH enterprises treat tooling choice as a compliance decision. Customers like Commerzbank and Continental operate in environments where every employee-facing video has to survive scrutiny from legal, data protection, and the works council. Choosing AI video with consent, brand control, and disclosure built in is what lets those teams move at all.
What should you do before 2 August 2026?
Start with a 30 minute audit. Where does AI currently generate or alter video in your comms output? Which of those are synthetic presenters that need disclosure, and which are real employees who do not? Who owns editorial responsibility for each channel? Write the answers down; that document is the spine of your compliance story.
Then close the gaps with tooling, not headcount. The same Simpplr data that shows 52 percent weekly AI use also shows 29 percent of the people closest to AI are anxious about their job security (Simpplr, 2026). The reassuring message for your team is that governance done well protects their work; it does not replace it. Pick AI Video Agents that disclose, stay on brand, and carry the certifications, and the August deadline becomes a formality rather than a fire drill.
Key takeaways
The deadline is real and close. Article 50 transparency obligations enter into force on 2 August 2026, with a possible transitional period to 2 December 2026 for providers already on the market.
Two duties matter for comms. Synthetic video must be machine-readable-marked, and published deep fakes must be disclosed clearly at first exposure.
Real employee video is not a deep fake. Consent-based recording with Ella keeps much of internal comms outside the heaviest disclosure category.
Governance by design beats manual review. Brand Templates means outputs are controlled and traceable by construction, which is what a transparency rule rewards.
DACH raises the bar. DSGVO and Betriebsrat co-determination make consent and disclosure a works-council matter, so prepared teams win.
Where cofenster fits
cofenster is the AI video platform for enterprise communications. Our AI Video Agents (Theo, Ella, Milo) plus Brand Templates let Comms, HR, and Marketing teams produce on brand video at scale, with consent, brand control, and disclosure built into the workflow. We are ISO 27001:2022 certified, GDPR compliant, and EU AI Act compliant. If Article 50 is on your radar, book a live demo and we will show you how Commerzbank, Continental, and Hugo Boss produce governed video without slowing down. For the wider picture, see our podcast on enterprise internal comms trends.
Frequently asked questions
What are the penalties for ignoring the EU AI Act transparency rules?
The AI Act sets administrative fines for non-compliance, and transparency breaches sit in the tier that can reach significant percentages of global annual turnover. For most comms teams the bigger near-term risk is reputational: publishing undisclosed synthetic content of a real person erodes exactly the employee trust internal comms exists to build. Treating disclosure as a default, not an afterthought, protects both the balance sheet and the relationship.
Do we need to label video made with AI editing tools like auto-captions?
Generally no. Article 50(2) exempts AI that performs an assistive function for standard editing or does not substantially alter the input. Auto-captioning, cropping, and basic clean-up of a real recording fall under assistive editing. The obligation bites when AI generates the content itself, such as a synthetic presenter or a manipulated likeness, rather than tidying footage a real person produced.
Who is responsible for compliance, the AI vendor or our company?
Both, in different roles. The vendor is the provider and must mark synthetic outputs as machine-readable. Your company is usually the deployer and must disclose deep fakes you publish and inform people when they interact with AI. This is why vendor choice matters: a provider that marks outputs and documents its governance does half the work for you. Confirm certifications and disclosure features before you buy.
Photo by Proxyclick Visitor Management System on Unsplash
