The report
Google's Gemini API policy says logging is enabled by default for Interactions API calls and Generate Content calls in billing-enabled projects created on or after October 26, 2026. Logs may contain request and response content. Developers can turn logging off or adjust retention; the default automatic deletion marking period is 55 days, with 7, 14, 28 or 55 days selectable. Logs promoted into saved datasets do not follow that ordinary retention clock.
Stored is not the same as used for training. Google says that paid-project logs are not used to develop models by default; contributing a dataset is a separate opt-in, with different terms and privacy consequences. Product marketing and user assumptions should not erase that distinction. The existing contextual disclosure boundary now includes this provider-documented example. Nothing in the documentation alone establishes a leak, violation or exposed customer project.
Settings that need separate verification
The project creation date, whether billing was enabled, and the API surface determine which default applies. A new paid Generate Content project created on or after October 26 is not equivalent to an older project, and the Interactions API has a distinct default. Applications that use an OpenAI-compatible endpoint should check how that request is treated under the same project policy rather than infer privacy from the client library name. A request-level store: false option and project-level settings are operational controls, but saved datasets have separate retention semantics. None of these storage choices alone grants a training license; participating in a contribution program requires its own affirmative terms. An accurate incident review must distinguish what was stored, who could access it, how long it remained, and whether any secondary training use was authorized.
A retention setting is not a disclosure finding
Google's policy distinguishes developer-owned logs from abuse-monitoring data, and ordinary paid-project storage from datasets voluntarily shared for model improvement. A 55-day default means records are marked for deletion after that period, not that an external observer has established their instant physical erasure. Datasets deliberately saved from logs have no automatic expiry specified by that default. This is a consent-and-lifecycle question for applications that process sensitive material: operators should check the API, billing state, project creation date and saved dataset policy separately. There is no reported breach or evidence of unauthorized model training in the cited document. See the contextual disclosure boundary, which distinguishes storing, seeing, using and sharing data.