Responsible AI, Safety & Risk for Architects

This module addresses the major security question for a live Claude system: what controls stop Claude from refusing a valid request, producing an unfair outcome, or taking an unapproved action. Safety is a set of controls spanning the request path, each with a blind spot the next must catch. The Architect places each one and decides what happens when it fails, because assuming Claude enforces a rule it was never given is the most common way a safety design will break.

Learning objectives
By the end of this module, you will be able to:

• Distinguish between what the model's training reduces and what your application layer must still enforce
• Place input screening, output screening, and tool-call authorization at the appropriate points in the request path and determine when to use model-based versus deterministic checks, so the system fails closed instead of failing open
• Identify where unequal outcomes can arise within a system and define the explanations required for users, regulators, and your own debugging team, so fairness and transparency are built into the design
• Route decisions to the appropriate reviewer/decision maker based on confidence, reversibility, and the cost of a wrong answer, so review effort is focused on the decisions that warrant them
• Map each compliance obligation to a named control, an owner, and an evidence artifact, so the architecture can be accurately audited