https://fluent2.microsoft.design/ai-harm

What this source teaches

  • Require an explicit harm model and severity-weighted test cases before any high-impact automation feature ships.
  • Distinguish AI harm categories (e.g., unfair outcomes, privacy, reliability, safety) as separate design and QA concerns, not a single 'risk' checkbox.
  • Link harm assessment outcomes to rule.ai.ux-failure-contract, rule.ai.ux-agency-contract, and rule.ai.ux-uncertainty-contract so each harm type maps to a concrete design response.

Where AwesomeDS applied it

Implementations and artifacts

No directly linked artifacts.

Caveats & anti-imitation

Adopt harm-model discipline and severity-weighted testing practices; do not import Microsoft product-specific harm classifications or brand materials.

Evidence: first-party-guidance · Verified 2026-07-16 · Cadence 60d