AI Disclosure and Transparency

A product page says the assistant is powered by AI. A help article says something slightly different. A contract says a third thing. When those three statements disagree, the fault lies in how each was drafted. This starter governs that drafting.

Disclosure language is written once and then quoted for years. A customer, a regulator or a journalist may hold you to a sentence that an assistant produced in a hurry. The governance lead who owns transparency statements needs a control path around that drafting work.

Before the model is called, injection screening and an identity check run in that order. Masking hides national identifier patterns inside whatever background material you paste in. A short list of blocked credential terms refuses a draft request that carries a secret, a private key or an access token.

The return path carries the weight here, because these words ship outward. A draft must reach a combined score of 0.85, with relevance held at 0.78 and accuracy at 0.80. Text under 120 characters or a single sentence is refused before a reviewer sees it. Grounding checks pull out the sources behind a claim and measure how well the text rests on them. In the shipped settings a weak source is flagged and the draft still comes back.

Decide what you must disclose before you use this. The starter does not read your obligations, track published versions or date a statement. Its review step is an empty block, so name an approver and switch it on before the first disclosure leaves the building. Audit marking notes that a request was evaluated, and storage and retention stay with your platform settings.