Ethical AI: Bias, Fairness, Transparency & Accountability

Beyond legal compliance, an architect addresses ethical risk. The exam framing treats these as testable, monitorable engineering properties -- not soft values bolted on after the system is built.

Bias and Fairness

Evaluate outputs for disparate treatment across groups, and include fairness cases in the eval set (the same evaluation infrastructure used for accuracy and safety testing). Fairness is not assessed separately from the rest of quality assurance -- it's another dimension of it.

Transparency

• Be clear with users that they're interacting with AI
• Cite sources for grounded claims
• Document how the system makes decisions

Accountability and Diligence

The AI Fluency framing -- delegation with description, discernment, and diligence -- puts a human accountable for AI-assisted outputs. The discipline is to verify rather than assume: a human delegates a task to the AI (with a clear description of what's being delegated), applies discernment in evaluating the result, and exercises diligence in checking it before it's relied upon. Accountability doesn't disappear just because a task was delegated to an agent.

Usage Policy (AUP)

The design must comply with Anthropic's universal standards and any high-risk-use requirements; agentic and MCP use carry additional responsibilities beyond the baseline policy, reflecting the larger blast radius of autonomous, tool-using systems.

Common exam traps

• Treating ethics as separate from engineering. Bias, fairness, and transparency are testable, monitorable properties that belong in the eval, the guardrails, and the documentation -- not a checklist item handled outside the technical design.