Databricks AI & Foundation Model APIs
Databricks says it does not train generative foundation models on data, prompts, or responses submitted to Databricks AI assistive features, and partner-powered features use zero-retention endpoints.
Customer prompts and responses are not used to train generative foundation models or to generate suggestions shown to other customers. Partner model providers are stated not to store prompts or responses for those assistive features.
No opt-out is required for the stated training baseline. Teams still choose which AI assistive features, Foundation Model APIs, and serving modes are enabled.
Databricks-hosted model processing is described as remaining in the Databricks security perimeter for those paths, while pay-per-token Foundation Model APIs can involve designated-service geography rules that require region review.
Databricks-hosted model paths may temporarily retain inputs and outputs for up to 30 days for abuse detection. Partner-powered assistive features are described as zero retention at the partner.
The reviewed trust pages focus on training restrictions, partner retention, and service architecture rather than one general human-review statement.
Foundational reference pages for teams adopting, deploying, and operating production AI with inspectable controls.