A banner is not evidence.
Most teams satisfy an AI disclosure obligation with a line of copy in a template. Then someone asks what it said in March, and on which channel, and whether it appeared before the first message — and there is nothing to point at. The AI Nutrition Label composes the notice as a content-addressed, signed artifact, keeps a delivery ledger that records whether it arrived in time, and lets anyone holding the id check it without an account.
Five elements, in every language you serve.
One of them is mandatory and the rest are yours to choose. The AI Nutrition Label will not compose a notice that omits the first — in any locale it carries.
It is a machine
The statutory core, and the one element composition refuses to publish without. A notice that never says the thing is artificial is not a disclosure, whatever else it contains.
On whose behalf
Who the system acts for. A person deciding whether to trust an answer is really deciding whether to trust whoever stands behind it.
What it can and cannot do
Capabilities and limits, stated before the conversation rather than discovered during it.
How to reach a human
The fallback, in the notice itself — not three menus deep.
Where to learn more
A durable pointer for the reader who wants the full picture.
A disclosure whose wording can drift afterwards evidences nothing.
The id is the notice
A label's identifier is the SHA-256 of its own canonical bytes. Reword it and you get a new notice with a new id — never a quietly changed one. That is what makes "we showed them this" checkable a year later by someone holding only the id.
Late stays late
Every delivery records how many assistant turns had already gone out when the notice appeared. Zero means it came first. The number is recorded, never assumed, so a late disclosure reads as late instead of being rounded up to "delivered".
Unknown is not covered
Coverage answers covered, not covered, or unknown — three states, not two. If we cannot read the answer we say so. An unreadable store is an infrastructure fault, not a clean bill of health, and it is never reported as one.
Refuses rather than weakens
Composition rejects a notice missing its core element, one that carries an element as an empty string, or one registered only to the machine-to-machine channel where no person is ever reached. A weak notice that looks complete is worse than none.
Anyone can check it
Resolve and verify take no account and no API key. A regulator, a counterparty, or the person who saw the notice can fetch the bytes and check the signature against the published key, with no trust in us at all.
Every locale, or it tells you
A notice can carry any number of locales. When a reader's language is missing, the label serves a fallback and RECORDS that it was a fallback — a coverage weakness worth seeing, not an error worth hiding.
We compose and evidence the notice. We do not decide whether you owe one.
Whether a particular system triggers a disclosure duty in your jurisdiction is a legal question about your deployment, and the AI Nutrition Label does not answer it — that is your counsel's call, not a product feature. What it does is make the answer actionable once you have it: compose the notice, deliver it, and hold a record that stands up when someone asks what was shown and when.
Compose it once. Prove it forever.
The AI Nutrition Label is a whole product on its own — and free. Combine it with the rest of Trustvice when you want to — through a workflow step you configure, never a dependency you inherit.
Trustvice 360Declare as regras. Deixe cada execução ser verificada.
Agentes, evidências, governação, comércio e divulgação estão no Trustvice 360. Registe-se grátis e abra todas as suítes a partir de um só painel.