Since 2 August 2026 the transparency rules of the EU AI Act apply. Anyone who builds generative AI must mark its output so machines can detect it, and anyone who publishes a deepfake must label it. At the same time deepfake attacks keep climbing: one tracker verified 821 of them in the first half of 2026 alone. This article explains what Article 50 requires, what Google, OpenAI, Anthropic, Mistral and Manus already mark, how watermarks and C2PA credentials work, and what a company in Croatia or the DACH region has to do now.
What Article 50 of the AI Act requires
Article 50 is the part of the AI Act that affects almost every company that uses AI in public, not only developers of high-risk systems. It has four obligations, split between providers (whoever builds or offers an AI system under their name) and deployers (whoever uses it professionally):
| Paragraph | Who | Obligation | Applies from |
|---|---|---|---|
| 50(1) AI interaction | Provider | People must know they are talking to an AI (chatbots, voice agents), unless it is obvious | 2 Aug 2026 |
| 50(2) Marking | Provider of generative AI | Audio, image, video and text output marked in a machine-readable format and detectable as AI-generated | 2 Aug 2026; 2 Dec 2026 for systems already on the market* |
| 50(3) Emotion recognition, biometric categorisation | Deployer | Inform the people exposed to it | 2 Aug 2026 |
| 50(4) Deepfakes and public-interest text | Deployer | Visible label on deepfakes and on AI text published to inform the public, unless a human reviewed it and an editor is responsible | 2 Aug 2026 |
* The four-month extension comes from the provisional AI Omnibus agreement of May 2026 and covers only the marking duty for generative AI systems placed on the market before 2 August 2026. Everything else has applied without delay. Breaches fall under the general penalty tier of Article 99: up to 15 million euros or 3% of worldwide annual turnover, whichever is higher.
The trap for ordinary companies is the provider role. If you build a customer chatbot on the Claude or OpenAI API and offer it under your own brand, you are the provider of that AI system for Article 50(1): your bot has to say that it is an AI. If you generate a marketing image showing a real-looking person, place or event, you are the deployer of a deepfake under Article 50(4). We covered the rest of the regulation in our AI Act compliance guide.
The Code of Practice: two layers of marks and EU icons
On 10 June 2026 the European Commission published the final Code of Practice on marking and labelling AI-generated content. It is voluntary, but it is the practical template for what authorities will expect. Signatories may use it to show compliance; everyone else has to prove by other means that their measures are adequate.
- Section 1, providers. No single technique is enough, so signatories combine at least two machine-readable layers where needed, typically signed metadata (C2PA) plus an invisible watermark. They must also offer detection so that users, authorities, researchers and media can check content.
- Section 2, deployers. Deepfakes and AI-written public-interest text get a clear label at the first exposure. The Commission published three free EU icons: a basic icon, "fully AI-generated" and "partially AI-modified". Using the icons is optional; labelling is not. For audio-only content an audio disclaimer replaces the visual label.
- Exceptions. Evidently artistic, satirical or fictional works need only a light-touch disclosure that does not spoil the work, and text that a human reviewed and an editor took responsibility for needs no label.
Source: European Commission, "Strong backing for the Code of Practice on Transparency of AI-generated Content" (31 July 2026, updated 24 September 2026). About 190 organisations in total; many signed both sections.
Section 1 reads like a list of the AI industry: Anthropic, OpenAI, Google, Microsoft, Mistral, Meta, Aleph Alpha, Black Forest Labs, Cohere and Synthesia, among others. Section 2 shows that the deployer side is not only media: Getty Images, Lufthansa, Lenovo, Iberdrola, Bulgari and Fastweb signed it. According to the Commission, about half of all signatories are small or young companies.
How the marks work: watermarks, C2PA and detection
Three technologies do the work, and they complement each other because each fails in a different way:
| Technique | How it works | Survives | Weak spot |
|---|---|---|---|
| C2PA Content Credentials | Cryptographically signed metadata in the file: which tool created or edited it and when | Copying the file unchanged; tampering is detectable | Lost on screenshots, re-encoding, many upload pipelines and image optimisers |
| Invisible image/video/audio watermark (e.g. SynthID) | A pattern woven into the pixels or the audio signal | Compression, cropping, resizing, screenshots | Detectable only with the provider's detector; heavy edits weaken it |
| Statistical text watermark (e.g. SynthID-Text) | A key nudges the model's word choice; detection checks the pattern | Light editing, copy and paste | Short texts, factual passages, code and full rewrites carry little or no signal |
What Google, OpenAI, Anthropic, Mistral and Manus mark
Sources: provider announcements and help pages, European Commission signatory list, press coverage of Google I/O 2026 and OpenAI (May 2026). "Partial" = preview, limited access or reported but not confirmed by the provider; "Not found" = no public documentation found, which does not prove the feature is missing; "n/a" = Claude does not generate images or video.
- Google started the trend with SynthID in 2023 and says that around 100 billion images and videos and about 60,000 years of audio carry the watermark (its own figure, I/O May 2026). Gemini text uses SynthID-Text, and Google adds C2PA credentials. In May 2026 OpenAI, Kakao and ElevenLabs adopted SynthID, building on an earlier collaboration with NVIDIA, and Apple announced it for iOS 27. According to reports from October, the SynthID Detector is now publicly available in English.
- OpenAI has attached C2PA metadata to its images since 2024. In May 2026 it joined the C2PA steering committee and added SynthID to all images from ChatGPT, the API and Codex, together with a preview of a public verification tool. Whether ChatGPT text carries a watermark is not documented in an official announcement we could verify.
- Anthropic published the most detailed description. Since 14 August 2026 Claude text carries a statistical watermark based on SynthID-Text: nothing is added to the text, it contains no user or company data, and in a side-by-side study human raters saw no quality difference. It applies worldwide, in the API, the Claude apps, Claude Code, Cowork and on the cloud platforms. Files (PNG, JPEG, SVG) get signed C2PA credentials, which the free Claude Content Checker can read. The detection API for text is in private preview for regulators, media, fact-checkers, researchers and companies with compliance obligations.
- Mistral signed Section 1 of the Code, so it has committed to marking. As of October we found no public technical description of how outputs from Mistral Vibe (formerly Le Chat) or the API are marked. Ask for it in your contract if you use Mistral for content.
- Manus is an agent platform built on models from other labs. We found neither a signature in the Commission's published examples nor documentation on marking. For companies this matters less than it seems: if you publish what an agent produces, the deployer duties of Article 50(4) are yours anyway.
Deepfakes: attacks are growing faster than the rules
The transparency rules come at a time when deepfakes have stopped being a curiosity. Resemble AI, which collects verified incidents from news reports, counted 1,567 incidents in 2025 and 821 in the first half of 2026, a pace that would slightly exceed last year.
Source: Resemble AI, 2025 Deepfake Threat Report (March 2026) and H1 2026 Deepfake Threat Report. The third bar doubles the H1 figure and is an extrapolation, not a measurement. Only incidents covered by the news are counted, so the real number is far higher.
The money is hard to pin down because most victims never report a figure. Resemble AI documented over 1.28 billion dollars in reported losses for 2025, and more than 80% of incidents disclosed no damage at all. The FBI received 22,364 complaints mentioning AI in 2025, with 893 million dollars in losses. Entrust found that deepfakes were one in five biometric fraud attempts, and in an iProov test with 2,000 people only 0.1% correctly classified every real and fake sample. The best-known corporate case is still Arup, where an employee transferred 25 million dollars after a video call with deepfaked colleagues.
Source: Resemble AI threat reports. Reach is the potential audience of the outlets that reported or spread an incident, an upper bound rather than confirmed views. In H1 2026, Grok accounted for 87% of the countable synthetic files.
Two trends stand out. Deepfakes that damage a brand or a person's reputation already reached more people in six months of 2026 than in all of 2025, and political disinformation grew by more than half; 22 political incidents each reached over a billion potential viewers. Corporate fraud, by contrast, almost disappears from the reach statistics: only 15 incidents in H1 2026, nearly all in trade media. That does not mean it is rare, only that companies do not talk about it.
What companies in Croatia and the DACH region should do now
- Inventory where you publish AI content. Marketing images and video, AI voice in ads or phone lines, chatbots on the website, AI-written press releases and blog posts. For each, note who is the provider and who is the deployer.
- Make chatbots and voice agents say they are AI. One sentence at the start of the conversation is enough, but it must be there in every language you offer.
- Label deepfakes and keep editorial responsibility for text. Use the EU icons or a clear text label on realistic synthetic people, places and events. For AI-assisted texts on public matters, document the human review and name the responsible editor, otherwise they need a label.
- Do not strip the marks. Many CMS, DAM systems and image optimisers remove metadata by default, which deletes the C2PA credentials. Check your pipeline and keep them.
- Ask vendors in writing. Which layers do they mark, from which date, in which products, and how can you or an authority check it? Add it to your AI procurement checklist.
- Protect payments against deepfakes. Call-back verification on a known number for every payment change, a second approver above a threshold, and a code word for urgent requests by "the CEO". Train finance and HR teams with real examples; see our AI cybersecurity guide.
Frequently asked questions
What does Article 50 of the EU AI Act require?
Chatbots must disclose that they are AI, generative AI output must carry machine-readable marks, and deepfakes and AI-written public-interest text must be labelled visibly. It applies since 2 August 2026.
Is there a grace period for watermarking under the AI Act?
Only for marking by generative AI systems already on the market before 2 August 2026: until 2 December 2026 (provisional AI Omnibus agreement). Deepfake labelling has no delay.
Does Claude watermark its text?
Yes, since August 2026: a SynthID-Text-based statistical watermark in the text and C2PA credentials in generated files. Text detection is in private preview.
Who signed the EU Code of Practice on AI-generated content?
About 190 organisations: 95 for the provider section and 192 for the deployer section, including Anthropic, OpenAI, Google, Microsoft, Mistral and Meta.
Can a watermark prove that content is AI-generated or human-made?
No. A found mark is a strong signal; a missing mark proves nothing. Keep your own production records.
Is your AI content ready for Article 50?
We help companies in Croatia and the DACH region map where they publish AI content, set up chatbot disclosures, labels and editorial review, and choose AI tools whose marking holds up to an audit. As a Claude Certified Architect from Zagreb, we start with your processes, not with a slide deck.
Talk to an AI consultant