Last updated: not yet published.
Only the specific query text being measured, or the specific content an analysis feature is run against (e.g. a crawled page excerpt, a piece of customer-voice text after PII redaction — see /legal/privacy) — sent via that provider’s standard API, using credentials the organization itself supplies or an organization-level key. No client is ever sent to a provider without a real, configured API key present.
As a general rule, every major AI provider’s API tier (as distinct from that provider’s free consumer product) publicly states that API inputs and outputs are not used to train its models by default, unless a customer separately opts in. This is the provider’s own commitment under its own terms, not something SignalOS’s code enforces or can verify at request time — SignalOS has no technical visibility into what a provider does with a request after sending it.
Before this is stated to a customer as a confirmed fact rather than a general expectation, someone at this company should:
A signed, provider-specific no-training commitment (where a provider offers one), and a real process for re-verifying this disclosure when a provider changes its terms, are not in place — this page states the general expectation, not a confirmed guarantee.