🇬🇧 This guide is also available in Romanian. / 🇷🇴 Preferi să citești în română? [Click aici pentru versiunea în limba română].
Starting August 2, 2026, the European Union’s Artificial Intelligence Act (AI Act) comes into force across all member states, including Romania. For online stores, marketplaces, and agencies, the most significant shift involves transparency requirements for content generated or modified by AI—ranging from “polished” product photos to descriptions written by an LLM or even support chatbots.
The European Commission recently published the “Code of Practice on Transparency of AI-Generated Content.” This voluntary document is highly practical: by adopting and applying it, businesses can more easily demonstrate compliance with Article 50 of the AI Act. We have reviewed it from the perspective of an e-commerce business and distilled it into clear technical requirements and simple steps to help you align with these best practices.
What is in this code and why does it matter?
The code has two main sections, addressed to different players in the ecosystem, but both are relevant if you are an online retailer.
The first section is for providers of generative AI systems—the companies developing tools for image, text, or video generation, such as those powering Adobe Firefly, Midjourney, Nano Banana, or other GPT-based tools. They are obligated to technically encode AI-generated or modified content using digitally signed metadata, invisible watermarks, or both, so that the artificial origin can be automatically “read” by digital systems.
The second section is for those who deploy or use these systems—organizations using these tools in their daily operations. This category includes your store, marketing agency, or any brand that generates hyper-realistic AI images or public texts.
For example, if you ask an agency for a banner featuring a product worn by a fully AI-generated model, or if you publish a blog post written exclusively by AI without human review, you fall under these rules. It doesn’t matter if the technology is yours or not; what matters is whether you use it.
When does this happen?
Industry news and European portal updates have already confirmed the key deadlines. As of this writing, transparency rules under Article 50 become mandatory for AI systems placed on the market starting August 2, 2026. Generative systems already on the market before August have a grace period until December 2, 2026, to implement these technical marking functions.
In Romania, ANCOM has been proposed as the national supervisory authority, but the implementation law—which will establish the local control mechanism—is still being drafted. However, this does not suspend your obligations: the European regulation applies directly, regardless of the status of local legislation. Simply put, the transparency obligation is a legal requirement, even if local enforcement is still organizing.
What online stores need to do
AI-generated or modified product images
If you use AI for lifestyle shots, background changes, dressing a virtual model, or generating color variants that weren’t actually photographed, these materials may fall under the category of hyper-realistic simulations—technically defined as “deep fakes” under Article 3(60). Basically, any content that could seem authentic to a visitor with no reason to doubt its origin.
The Code of Practice recommends a standardized “AI” icon, provided for free by the European Commission, in two versions: “AI GENERATED” for elements created from scratch, and “AI MODIFIED” for real images that have been substantially altered. The icon must be placed visibly, for example in the corners of the image, and cannot be obscured by other graphic elements.
Artistic, creative, or satirical works aren’t completely exempt but have more flexible placement rules: the label can appear in the material’s description, end credits, or presentation brochure, rather than being “overlaid” directly on the image, so it doesn’t ruin the work’s experience. However, standard catalogs in an online store do not qualify as works of art; for these, the standard labeling rules above apply.
Product descriptions and blog articles
For text, there is an important distinction: labeling is only necessary if the content is published to inform the public on matters of general interest and is published exactly as generated, without human editorial control.
Technical product descriptions or short SEO texts do not fall under this rule. Guides, market analyses, or educational materials, however, may be targeted if they are generated exclusively by AI and published without anyone taking responsibility for them.
What does this mean in practice? The internal process of human validation is what saves you from labeling. If someone on your team reads the generated text, edits it, and takes ownership of the final version, the text no longer needs to be marked as AI-generated.
Chatbots and virtual assistants
These tools are not mentioned in the Code of Practice covering paragraphs (2), (4), and (5) of Article 50 of the AI Act, but they are directly addressed in paragraph (1) of the same article, with the same deadline: August 2, 2026. According to the document, online stores using chatbots or virtual assistants must inform users from the start that they are interacting with a software system, not a human (unless this is obvious).
AI-generated video or audio ads
For AI-generated video banners or social media campaigns featuring realistically rendered people, places, or events, the same strict rules apply: a visible label at the beginning of the material, repeated after commercial breaks, and for audio-only content, a clear audible announcement in the first few seconds.
Technical requirements for in-house developers
For agencies building internal AI tools, Section 1 of the Code establishes four standards for marking systems that must be met simultaneously:
- Effectiveness (the end-user easily notices and understands the mark),
- Reliability (a very low error rate),
- Robustness (the mark resists cropping, resizing, screenshots, compression),
- Interoperability (can be identified by third-party systems).
The recommended approach has two directions: digitally secured metadata (invisible but technically verifiable) plus an imperceptible watermark integrated into the file. For simple texts that cannot carry metadata, an invisible watermark is considered sufficient. Very short texts, under approximately 200 tokens, are exempt from watermarking because current technology cannot guarantee reliable marking on such a small volume of words.
Beyond compliance: Customer trust
Transparency and honesty regarding AI content have become, independent of the law, an increasingly important factor for shoppers. In an online store, visual trust in a product directly influences conversion rates. This is especially true in fields with complex products, where digitally generated color variants or finishes might look visibly different from reality if not correctly identified.
A store that communicates openly—for example, by stating an image is an AI rendering for a color variant not yet available in the photo shoot—protects its reputation and reduces, at least theoretically, returns caused by discrepancies between online visuals and delivered items. It is no coincidence that the AI Act emphasizes the risk of such discrepancies as one of the central reasons for introducing labeling obligations.
Steps for alignment
- Inventory the AI tools used for images, descriptions, customer support, and video materials.
- Evaluate content: identify materials that could be confused with real photos or informative texts published without human validation.
- Implement an editorial review process: a team member reads, edits, and takes ownership of the final version of AI-written texts.
- Suggestive icons: integrate them proactively where they make sense, especially in promotional banners and substantially modified images.
- Document decisions: even informally, an internal log is the best proof of good faith when control mechanisms become fully functional.
Closing thoughts
The AI Act does not discourage the use of artificial intelligence in e-commerce. Image generation, virtual assistants, or optimized texts remain perfectly legal tools. What changes is the obligation not to hide the artificial origin of content, especially when the lack of this information could be misleading.
For an online store or agency, the transparency requested by the AI Act can be seen in a different light: as another reason to communicate better with your customers, rather than just another unwanted bureaucratic burden.
Send us a message via the form below, and we’ll help you determine exactly what your store needs to do to follow the new rules.
Note: This article is aimed at online store owners and e-commerce agencies using AI for product images, descriptions, chatbots, or ads. It is for informational purposes only and aims to bring a current topic closer to the e-commerce community. The information provided does not constitute legal advice; for compliance decisions, we recommend consulting a legal advisor or obtaining complete information from official sources.
For additional technical details about the AI Act, you can consult resources at artificialintelligenceact.eu.
The AI Act regulatory framework is in its initial implementation phase; some deadlines or procedures (including national sanction laws in Romania) may change. We recommend consulting information directly from the European Commission’s dedicated AI Act website and any official statements from ANCOM or other Romanian authorities before making business decisions based on this article.