Hello,
Previously, Google had offered Gemini inference with Zero Data Retention over Vertex, made available via OpenRouter. Now it seems that instead there is a carveout in which prompts may be logged via a separate “safety” classifier. However, as a user, a safety classifier which may or may not retain arbitrary inputs is functionally no different from logging. It entails that all of the promises that one can then make downstream of ZDR can’t be consistently stated - “Your inputs are not retained, unless they are” is not very different from “Your inputs are retained, and may or may not be reviewed”, if one needs to be able to provide guarantees about this.
I would like to politely request the restoration of proper ZDR as it was under the previous terms. I was directed here from the Gemini API forum. If this isn’t the right place, could this request be forwarded to the appropriate team? Thank you for your time.