Tools for Transparency and Explainability

Transparency only works if the right information is written down and easy to find. AWS offers several ways to document and check models: SageMaker Model Cards for models you build, AI Service Cards for AWS's own AI services, Bedrock evaluation jobs for evidence of how a model performs, and model licenses and open weights for what you're allowed to inspect and do. This lesson matches each one to the question it answers, which is how the exam asks about them.

SageMaker Model Cards

Amazon SageMaker Model Cards document a model you own in one place. A card records the model's intended uses (and uses it isn't meant for), a risk rating of unknown, low, medium or high, training details, evaluation results, and notes such as ethical considerations and caveats. Cards serve two audiences: people deciding whether and how to use the model, and auditors checking how it was built and tested.

Two features matter for governance. You can export a card as a PDF to share with stakeholders who don't use AWS. And every edit, apart from an approval status change, creates a new version, so there's a lasting record of how the model and its documentation changed. Model Cards are also integrated with SageMaker Model Registry, so audit details can be attached to registered model versions.

AWS AI Service Cards

When you use a managed AWS AI service, such as Amazon Polly or Amazon Bedrock Guardrails, you didn't build the model, so you can't write its model card. Instead, AWS publishes AI Service Cards. Each one explains the service's intended use cases and limitations, the responsible AI choices AWS made in its design, and best practices for deploying it and getting good performance. They're AWS's own transparency documentation, and AWS still recommends testing each service on your own data.

Evaluation Results as Evidence

A claim that a model is accurate and safe is only as good as the testing behind it. Amazon Bedrock Model Evaluations produce that evidence: automatic jobs score models on metrics such as accuracy, robustness and toxicity, and human or judge-model jobs rate qualities that need judgment. The results give a review board something concrete to approve, and they can be recorded in the model's documentation so that later readers see how it was tested.

Open Models, Data and Licenses

How much you can see inside a model depends on where it comes from. An open-weight model, such as the publicly available models in Amazon SageMaker JumpStart, can be deployed, fine-tuned and tested in your own account, and each comes with a license you check. A proprietary model is used through an API, so you rely on the provider's documentation instead.

Either way, every model comes with a license. JumpStart models are released under the terms their providers set, from permissive open-source licenses to ones with usage restrictions, and some require you to accept an end-user license agreement first. AWS states plainly that reviewing and following those terms is your responsibility. Information about the training data, where a provider shares it, is part of the same picture: it tells you what the model has and hasn't seen.

For explaining individual predictions of custom models, AWS's tool has been SageMaker Clarify. It's now closed to new customers, though existing customers can still use it.

On the Exam