HomeDisclosure under Art. 50 AI Act
AI transparency
A large share of the images on this site was produced by generative models; the films, by contrast, are drawn from code, and the articles are written with AI assistance. This page says which parts that covers, which parts it explicitly does not, and how you can tell while you are reading. The basis is Article 50 of Regulation (EU) 2024/1689 on artificial intelligence, whose transparency obligations have applied since 2 August 2026.
What is marked
Every image and every film a generative model produced carries the mark “AI-generated” in the top-right corner of its frame. It sits over the picture rather than under it, and it is visible without scrolling and without a click — Article 50(5) requires disclosure at the time of first exposure at the latest.
For screen readers the same mark carries the full sentence “AI-generated image” as its accessible name. It is therefore not decorative, unlike every other overlay on a frame here.
Whether a mark is set is decided in code rather than by hand. The default is to disclose: a frame whose origin is not explicitly recorded as a recording or a drawing gets the mark. A forgotten image is over-marked rather than under-marked, and that is the intended direction.
Images and films
Every photographic-looking picture on this site is generated, not taken. That covers the hall and hangar motifs, the city bands on the location pages, the still lifes on the service and case-study pages, and the environment shots in the articles.
They show real objects, places and situations in a register that can be read as photography. They therefore meet the definition of a deepfake in Article 3(60) of the AI Act, and the disclosure duty in Article 50(4) applies — regardless of the fact that none of them shows another person and none asserts an event that did not happen.
This expressly does not cover the films. They are drawn from code and are not model output; "What is not generated" below says why that is a difference and not a matter of interpretation.
Pictures of me
The portraits on this site are produced from a model trained solely on approved, current photographs of myself. They show me, they are not photographs, and they carry the same mark as any other generated image.
No other real person is generated — not clients, not staff, not third parties. Consent for the person depicted exists because that person and the operator of this site are the same one.
What is not generated
Three classes of material are explicitly not model output and therefore carry no mark:
- Screen recordings and screenshots. Interfaces, terminals, measurements, deployments and error messages are recorded, never generated. An invented interface would be an invented piece of evidence.
- Diagrams, schematics and the films. They are drawn from code — deterministic programs that emit the same pixels on every run. That this code was written with AI assistance does not make its output AI-generated content: Article 50(4) attaches to content generated or manipulated by an AI system, and an AI system under Article 3(1) is one that infers from its input how to generate its output. A renderer with no model in it infers nothing.
- The figures, timescales and outcomes in the case studies. They come from the projects, not from a model.
Text
The articles are written with AI assistance and are reviewed and released on their substance before publication. Editorial responsibility for every published text rests with Damjan Savić, at the address and contact given in the legal notice.
Article 50(4) subparagraph 2 of the AI Act requires disclosure for AI-generated text published in order to inform the public on matters of public interest, and exempts text that underwent human review and for which someone assumes editorial responsibility. The articles here are technical writing about software rather than reporting on politics, public administration, the judiciary or fundamental rights; the exemption applies in addition. This statement is therefore voluntary.
The legal texts — legal notice, privacy policy, terms and this page — are written against the running system and checked against it. They have not been reviewed by a lawyer, and that limitation is stated here rather than in small print.
Machine-readable marking
Article 50(2) of the AI Act obliges providers of generative systems to mark their outputs in a machine-readable format. That duty falls on the model providers, not on the operator of this site.
What actually happens still belongs here. No machine-readable marking of origin is present on the image files this site delivers: no C2PA manifest, no signature, in any file. That is measured rather than inferred — every image file under public/media is read before release, and the one file that carries a metadata block at all holds an editing history from Photoshop and no statement about origin.
The reason is explicitly not that something is being removed here. The generators used embed no provenance signature, so there is none that could be lost. What would happen to such a signature depends on the path: images served through the image optimiser are re-encoded and would lose it; video posters and anything served straight from the directory would keep it intact.
The visible mark on the frame is the disclosure; it does not replace a signature and is not offered here as one. If a provider starts shipping a provenance signature, this paragraph is wrong. It is therefore not believed but re-measured: a check in the project reads every media file and reports exactly that case, so the statement here can be corrected.
There is a machine-readable statement all the same, and it is deliberately not in the file: the portrait this site names in its structured data as the picture of the person carries the IPTC property digitalSourceType there, with the value trainedAlgorithmicMedia. It sits in the page source rather than in the image data, which is why it survives the re-encoding. It is a statement by the operator and not cryptographically verified provenance — that difference is why the paragraph above is still true.
Systems used
This list names what was actually used. It is not a statement about what could be used.
- Image and video generation: Higgsfield, including an identity model trained on my own photographs.
- Text: Claude by Anthropic, for drafting, editing and translation.
- Code: Claude Code, for building this website.
- Visitor search: a local BM25 index; no model provider receives search queries.
AI literacy
Article 4 of the AI Act requires providers and deployers to take measures that support the development of AI literacy among people operating these systems on their behalf; it does not require a guaranteed individual level. This business has no employees. The systems are operated by the responsible person, whose profession is building systems of this kind.
Questions and corrections
If you notice an image or a text on this site that is marked wrongly or not at all, write to info@damjan-savic.com. An informal message is enough, and a correction follows promptly.