The retailers’ request

European retailers would like the law to make an exception for them. According to a Reuters report, a retail trade association has called for AI-generated advertising to be carved out of the European Union’s incoming transparency rules, which would require companies to label commercial content produced by artificial intelligence before those rules start to bite.

The request comes as the EU AI Act fine-tunes its enforcement mechanisms, and it reveals how much the advertising industry fears significant operational impact on a practice already widespread: the automatic creation of ads, product descriptions, and personalized creatives using language models. The sought-after exemption would apply only to commercial communications, leaving intact the obligations for deepfakes or other forms deemed potentially deceptive.

The new transparency obligations under the AI Act

The transparency provisions in the AI Act state that any content generated or manipulated with the help of artificial intelligence, when intended to inform the public on matters of public interest, must be clearly labeled as such. For advertising, the requirement is even stricter: the end user must be able to unambiguously recognize that they are viewing an artificially produced ad.

The goal is to counter phenomena such as disinformation and deepfakes, but the line between what is deceptive and what is merely automated remains thin. In retail, where much of the promotional material is already generated with generative AI tools, the prospect of having to affix a label to every banner, video, or personalized text creates uncertainty, especially when the generation takes place through self-hosted pipelines rather than via external cloud services.

What changes for LLM-based advertising pipelines

For retailers that have chosen to run LLM models on-premise (self-hosted) – often to maintain full control over data and reduce dependence on third-party vendors – the labeling obligation introduces an additional layer of complexity. Running inference on in-house hardware does not exempt the output from transparency rules: the responsibility to indicate the artificial nature of the content falls on the producer, regardless of the underlying infrastructure.

This means that every piece generated by an LLM must be marked or accompanied by metadata attesting to its synthetic origin. To do this systematically, companies need to integrate auditing and tagging tools into their advertising production pipelines, an activity that can slow time-to-market and requires a software architecture capable of tracking the genesis of each asset. Without an exemption, large-scale AI use for commercial purposes risks becoming a non-trivial compliance burden.

Transparency and local control: an AI-RADAR reading

This affair highlights a crucial issue for those evaluating on-premise LLM deployment not only as a lever for data sovereignty, but also as a tool for managing compliance. Local infrastructure allows the construction of more transparent generation flows, where labeling becomes part of the content lifecycle from the design phase. At the same time, however, a potential exemption for the retail sector would signal that the regulatory framework recognizes the peculiarity of commercial communication, where the authenticity of the message is not threatened by the use of AI.

For technology decision-makers, the association’s request is a signal: the transparency game is played not only on the field of consumer protection, but also on that of competitiveness. Companies designing stacks for generative AI must therefore consider, beyond performance and total cost of ownership (TCO), the ability to demonstrate compliance with rules that could evolve rapidly. In this sense, a well-orchestrated self-hosted pipeline can also become an asset for reporting, but only if the architectural design embraces transparency as a system requirement, not as an afterthought.