Google Cloud Vertex AI

Google Cloud excludes training without permission, but caching, grounding, and session features introduce bounded retention.

Plan Scope

Generative AI on Vertex AI, including managed Google and partner models; grounding, caching, live sessions, stateful resources, and abuse-monitoring eligibility differ.

Prompt Retention

Stateless inference is not retained at rest by default, but Gemini uses project-isolated in-memory caching for up to 24 hours unless disabled. Grounding with Google Search or Maps stores prompts, context, and outputs for 30 days; session resumption caches content up to 24 hours.

Response Retention

Same feature-specific handling as prompts. Explicit context caches, batch files, tuned models, RAG stores, and other customer resources persist until their configured expiry or deletion.

Ordinary Logging

Some customers under Google Cloud Platform Terms are subject to prompt logging for abuse monitoring and can request an exception. Service, billing, and security metadata remains.

Model Training

Google contractually says it will not use customer data to train or fine-tune any AI/ML model without prior permission or instruction, including GA and pre-GA managed models.

Product Improvement

Customer content is excluded from general model improvement absent permission; grounding data may be used for debugging and testing the grounding systems during its 30-day retention.

Human Or Operator Access

Abuse and support access follows Google Cloud controls and Access Transparency where available; no universal operator-blind commitment applies to every free or feature path.

Subprocessors And Routing

Google Cloud and its subprocessors process data in the configured region or multi-region, while partner models and grounding services can add separate terms and locations.

Deletion Controls

Customers can disable in-memory caching, request an abuse-monitoring exception where eligible, avoid stored grounding/session features, and delete explicit caches and other Vertex resources.

Caveat

Vertex AI's no-training promise is strong, but achieving zero retention requires disabling or avoiding every documented cache, grounding, session, abuse, and stateful-storage path.