Microsoft Foundry direct models
Microsoft says prompts, outputs, embeddings, and training data for Foundry direct models are not used to train generative-AI foundation models without customer permission or instruction.
Customer inputs and outputs are not used to improve Microsoft or third-party products and are not used to train foundation models without explicit permission or instruction.
No opt-out is required for the stated training baseline. Deployment type, stateful features, feedback, and logging still require configuration review.
Microsoft says direct-model data is not available to other customers or model providers. Processing geography depends on regional, Data Zone, or Global deployment choices.
Base-model inference is described as stateless, while Responses, Assistants, stored completions, batch, files, and vector stores can persist or route data differently.
Abuse monitoring can select samples for automated and, when necessary, human review; qualifying managed customers may apply for modified monitoring.
Foundational reference pages for teams adopting, deploying, and operating production AI with inspectable controls.