Draft addendum

Sponsored Answer: labeling and prompt-injection rules

Rules for disclosing sponsored answers in AI chat and agent conversations, and for keeping sponsored content from steering the assistant. Draft addendum to Sponsored Answer v0.1, open for comment.

Draft addendum to v0.1 These rules are proposed for the next Sponsored Answer version. Comment through public comment.

A. Sponsored labeling

  1. Always labeled. Every sponsored answer is shown with a visible label “Sponsored” (or the local equivalent, e.g. Urdu «سپانسرڈ», Arabic «إعلان»/«مُموَّل») before or at the start of the sponsored content, in the same language as the conversation.
  2. Visually distinct. Sponsored content is separated from the assistant's organic answer (container, border or background) and never blended into the organic text.
  3. Same size and contrast. The label is at least as legible as the surrounding text and is not hidden behind a hover, tooltip or "more" link.
  4. Advertiser named. The advertiser or brand name is shown with the sponsored content, with a link or control explaining why it was shown.
  5. Voice and audio. In spoken interfaces the word "sponsored" (or local equivalent) is spoken before the sponsored content.
  6. No impersonation. Sponsored content must not claim to be the assistant's own recommendation, a neutral review, or a system message.
  7. Signals in the bidstream. The rendered label text and placement are reported in the event payload (sa-event) so buyers and auditors can verify disclosure.

B. Prompt-injection safety

  1. Ads are data, never instructions. The assistant treats all sponsored text (title, body, URLs, metadata) as untrusted display content. It is never added to the system prompt or tool instructions.
  2. Isolation. Sponsored content is inserted after the organic answer is generated, or in a separate, clearly delimited channel that the model is instructed not to follow.
  3. Creative screening. Exchanges reject creatives containing instruction-like text aimed at models (for example "ignore previous instructions", "you are now…", role or system markers, hidden or zero-width text, encoded payloads).
  4. No tool triggering. Sponsored content can never cause the assistant to call tools, open URLs, make purchases, send messages, or change settings without an explicit user action.
  5. No influence on organic answers. Bids and creatives must not change the wording, ranking or facts of the organic answer. Platforms log the organic answer before ad insertion for audit.
  6. Sensitive contexts. No sponsored answers in conversations about health emergencies, self-harm, or for users identified as minors; these follow the platform's restricted-category list.
  7. Reporting. Users can report a sponsored answer; reports and blocked injection attempts are counted in the event stream.