AI Transparency

Draft for legal review. Art. 50 of Regulation (EU) 2024/1689 applies from 2 August 2026; the Digital Omnibus deferred only the high-risk provisions (Annex III to 2 December 2027), not the transparency rules.

What the assistant is

An AI system that answers questions only from content the customer has indexed (retrieval-augmented generation). It does not browse the web and does not use your data to train models.

Disclosure to people using it (Art. 50(1))

Under Art. 50(1) the obligation to design the system so that people are told they are interacting with AI rests on the provider — that is us — not on the customer. The chat window states, at the latest at the first interaction, that the visitor is talking to an AI assistant.

As a customer you may restyle the widget, but keep the disclosure visible: the statutory exception only applies where it is already "obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use". A white-label chat widget is not obviously an AI to a reasonably circumspect shopper, so rely on the explicit label, not on the exception. Deployer duties under Art. 50(3)–(4) concern emotion recognition, deep fakes and text published to inform the public on matters of public interest — none of which a retail catalogue chat is.

Marking of generated content (Art. 50(2))

As the provider of a system that generates text, we mark its outputs in a machine-readable format, detectable as artificially generated, so far as this is technically feasible. API responses carry this flag together with the cited sources. The deployer duty to disclose generated text (Art. 50(4)) applies only to text published to inform the public on matters of public interest, so it does not arise for a shopper's one-to-one catalogue chat — the end consumer is not a "deployer".

How an answer is produced

  1. The question is rewritten into a search query.
  2. Search retrieves passages from the customer's index, applying the same access rules as search.
  3. Retrieved text is screened for instruction-like content (prompt-injection defence).
  4. A language model writes an answer only from those passages; without passages it says it could not find an answer.
  5. Each claim links to its source.

Limits

Answers can be incomplete or wrong. They are not professional (medical, legal, financial) advice. Customers should review answers for sensitive topics and can disable the assistant per index.

Models and providers

The model provider configured for the service is listed on Sub-processors.

Risk classification

Site search, recommendations and the assistant are not high-risk use cases under Annex III of the AI Act.

Questions: legal@aacsearch.com.