Transparency obligations enter force
On Sunday 2 August the first tranche of transparency obligations under the EU AI Act becomes applicable. The rules cover customer-facing chatbots, virtual assistants, image and audio output, deepfakes and AI-generated advertising visuals. Companies that breach the requirements face fines of up to 15 million euros or three percent of worldwide turnover, whichever is higher.
The Czech AI Association notes that the regulation does not demand a visible label on every output. Article 50 sets out five distinct obligations that apply to different actors in different situations and in different ways. For practical compliance two documents are decisive: the European Commission implementation rules on transparency obligations and the code of practice for transparency of AI-generated content.
What the rules require in practice
According to lawyer Jaroslav Kuby of the advisory firm Rödl, companies must clearly and promptly inform users that they are communicating with artificial intelligence via a chatbot or voice assistant. The same applies to AI agents that independently make reservations, place orders or handle correspondence; under the Commission guidance they must disclose not only that they are AI but also on whose behalf they act. Any AI-created or manipulated content that resembles real people or events, deepfakes, must carry machine-readable labelling. The rules also extend to advertising visuals and product photographs produced by AI where these could mislead about the actual appearance or properties of a product. The information must be understandable and placed where the interaction occurs.
High-risk deadlines shift but preparation cannot wait
The Digital Omnibus amends the original 2024 AI Act. Anton Arapov, Director of OpenSSL Corporation, points out that while labelling obligations now take effect, the rules for high-risk AI systems move to the end of 2027 and 2028. The ban on nudification applications and the centralisation of oversight of large providers through the European AI Office are likewise deferred. Arapov warns that labelling AI content is at heart a problem of signing and verification; labels function only when they are authentic, tamper-resistant and verifiable.
Verification not just labelling
Companies must immediately audit their AI systems and put in place both technical and process measures for labelling in line with the specific scenarios set out in Article 50. The risk of fines reaching three percent of turnover demands a systematic approach to compliance, not merely the addition of visual markers. Implementation of machine-readable watermarks and verifiable metadata becomes essential for deepfakes and advertising materials. The postponement of deadlines for high-risk systems does not create room to delay preparation; audits and risk analyses should already be under way.