Responsible AI
The accelerator does not yet automate Responsible AI (RAI) controls. This page lists what is and is not covered, and what is planned.
What you get today
| Area | Covered by | Notes |
|---|---|---|
| Infrastructure guardrails | policy_mode | Private access, no local auth, required tags. Not model behavior. |
| Network isolation, identity, logging | Scenarios 2, 3, 4 and 9 | Security posture, not fairness or safety. |
| Model fairness, explainability, error analysis | Nothing | Bring your own. |
| GenAI content filtering and safety evaluation | Nothing | Scenario 10 deploys AI Foundry without a content filter policy. |
Planned additions
The plan is an rai option, off by default, because RAI applies across scenarios. It is not implemented or verified live.
- Classic ML scenarios: a Responsible AI dashboard step in the training pipeline (fairness, explainability, error analysis) and a model-card template in the repo.
- Scenario 10 (GenAI with AI Foundry): a content filter policy on model deployments, then evaluation runs with the Foundry safety evaluators.
Until then
- Azure ML: Responsible AI dashboard and model interpretability.
- AI Foundry: Content filtering and Risk and safety evaluators.
- Add the checks to your own training pipeline under
src/. The generated CI/CD runs whatever the pipeline produces.
Scope
Tooling does not make a system responsible. The checks above support a review process. They do not replace one.