We use artificial intelligence in our day-to-day work: from avatar video and localization to visuals, copy, and analytics. With that comes our responsibility to let people know when they're looking at content a machine created. This statement describes our approach.
We are a signatory of the Code of Practice
Cognito.CZ, s.r.o. has joined Section 2 of the Code of Practice on transparency for AI-generated content, the part intended for deployers: those who use generative AI in their professional activity. We're included on the list of signatories published by the European Commission.
The Code was created under the auspices of the European Commission's AI Office, and in July 2026 the Commission and the AI Board assessed it as an adequate tool for demonstrating compliance with the transparency obligations under Article 50 of the EU AI Act (Regulation 2024/1689). We chose to sign it consciously and voluntarily, because we want clear, transparent rules for using AI.
This statement applies to the entire Cognito group.
What we label: primarily deep fakes
We don't label everything AI assisted with. We only label content a person could mistake for reality.
This matches the law. The obligation covers so-called deep fakes: image, audio, or video content that noticeably resembles existing people, objects, places, or events and could falsely appear authentic to a viewer. The European Commission's guidelines explicitly exclude content that is obviously unreal: anything that contradicts the laws of nature or biology and has no potential to mislead doesn't need labelling. Intent isn't examined — what matters is how the content can come across, not what we intended. And where content mixes informative and creative character, we go by the informative one.
We label
- AI avatars in video
- Photorealistic AI-generated people, even when they don't represent anyone specific
- The likeness or voice of a real person in a scene that could pass as real
- Real video where AI changes speech or voice (lipsync, dubbing)
- Photos where AI changes what the viewer takes to be an authentic record
- Visualizations that look like photos of a finished state or an existing place
We don't label
- Brand mascots, cartoon figures, and obviously fictional content
- A real person in an obviously impossible scene that can't be mistaken for a record
- Advertising stylization where the format makes it clear it's not documentary footage
- Post-production: color grading, retouching, sharpening, resizing
- Visualizations presented as visualizations, where the viewer expects a concept, not reality
- Marketing copy, product descriptions, and creative content
For visualizations, the product's actual properties and what will actually be delivered stay truthful either way. For text, the labelling obligation only applies to content published to inform the public on matters of public interest, which our written output isn't.
How we label
We treat compliance as two separate layers and work with both.
Machine-readable labelling — metadata and digital watermarks — is the responsibility of the AI tool providers we work with. When choosing vendors, we take into account how well they meet this obligation.
Visible and audible labelling for people is our own responsibility. We don't rely on a machine watermark the viewer never sees to handle it for us. We use the official icons issued by the European Commission's AI Office: the "AI" abbreviation with the word GENERATED for content created entirely by AI, or MODIFIED for real content that AI has altered.
The viewer sees the label right away, on first exposure, with no need to click through anything.
If the publishing platform itself can do this and displays a label directly alongside the content — not tucked away in a hidden menu — we rely on it and don't duplicate the label; two labels stacked on top of each other would only hurt readability. The Code recognizes such a solution as equivalent.
In every other case, we embed the label directly into the creative: at the start of a video and then repeated, burned into the image for photos, and as a short spoken announcement for audio. That is, whenever the platform hides the label in a menu, when we hand off the content as a file, or when it's going out across multiple channels. This way the label stays in place even where someone else publishes the content or it gets shared further.
We design the label to work for people with disabilities too. We keep contrast at WCAG 2.1 level AA and pay attention to readability for screen readers and for users with color vision deficiency. On a client's request we also prepare custom, on-brand labelling, always in line with accessibility and recognizability rules. For sensitive and regulated content (healthcare, finance, education, public administration) we use only the official icon.
The three icons we use
We take the icons in their original form from the European Commission's archive and never modify or recolor them. The Commission publishes an overview of the icons and the terms of their use, along with downloadable archives in SVG and PNG format. The Commission provides four color variants; we use the transparent ones only for clearly creative, artistic, and fictional works, where subtlety is appropriate. Everywhere else we choose the solid variant that fits the background.
What the label looks like in practice
Two of our own examples. In both, the label is burned directly into the video, in the top-right corner, clear of subtitles and other elements, so it stays in place even after downloading or further sharing. We pick the variant's color based on the background so the label stays readable.
A fully generated AI avatar — a news anchor who doesn't exist. AI created the entire image and the environment, hence GENERATED.
The background in the corner is dark, so we used the white variant.
Real footage of a real person whose speech AI changed. The original content is authentic; AI only modified it, hence MODIFIED.
The background in the corner is light, so we used the black variant.
Labelling doesn't replace consent
Labelling content is just one part of a responsible approach.
Wherever we work with the likeness or voice of a specific real person, we always secure their documented consent, regardless of the fact that we label the result. We make sure the label is never a cover for content that could mislead about the properties of a product, service, or place. A visualization must not promise what won't be delivered, and an AI "customer" must not recount an experience nobody actually had.
Our approach keeps evolving
The regulatory framework is new. As the interpretation gets refined through European Commission guidance, case law, and industry best practices, we'll adjust our approach accordingly. We revise our internal policy every six months, and ad hoc whenever the rules change. This includes training everyone who works with AI tools, including external collaborators, and keeping records of how we classify and label each output.
Where we're unsure, we'd rather label. Where a label would be pointless, we don't add one just for show.
