Phase 1: When AI Harms — Real-World Cases & the Responsible AI Landscape · 35 min · Python

Trustworthy AI Characteristics — What 'Responsible' Actually Means

Trustworthy is not a vibe — it's a checklist with tradeoffs.

Hiring signal: Responsible AI roles require the ability to evaluate an AI system against structured trustworthiness criteria and articulate the tradeoffs between characteristics. A candidate who can say 'we prioritized fairness over accuracy because the use case demands equalized odds, and here's the measured tradeoff' is demonstrating exactly the kind of structured reasoning RAI roles demand.

What you will learn

Enumerate and define the 7 NIST trustworthiness characteristics
Explain the key tradeoffs: fairness vs. accuracy, privacy vs. utility, explainability vs. performance
Articulate Chouldechova's theorem and why simultaneous fairness is mathematically impossible
Build a trustworthiness scoring rubric that evaluates an AI system against all 7 characteristics

The Problem

When someone says "we built a responsible AI system," what do they actually mean? Without a structured definition, "responsible" is a marketing word. NIST's AI RMF provides the answer: a trustworthy AI system demonstrates 7 specific characteristics. But here's the catch — you can't maximize all 7 simultaneously. Every real system involves tradeoffs, and a Responsible AI Engineer's job is to make those tradeoffs explicit, measured, and defensible.

The 7 Trustworthiness Characteristics

NIST AI RMF defines 7 characteristics that collectively constitute trustworthiness:

Characteristic	Definition	How You Measure It
Safe	The system does not cause physical, psychological, or societal harm	Incident rate, harm severity assessment, failure mode analysis
Secure	The system resists adversarial attacks and protects against unauthorized access	Penetration testing, adversarial robustness testing, OWASP LLM Top 10 compliance
Explainable	The system's outputs can be understood by relevant stakeholders	SHAP/LIME explanations, model cards, user-facing transparency
Privacy-Enhanced	The system protects training data and user information from leakage	Differential privacy (ε value), membership inference resistance, PII detection
Valid & Reliable	The system performs as intended across its operational domain	Accuracy, robustness, drift detection, out-of-distribution testing
Fair	The system produces equitable outcomes across demographic groups	Demographic parity, equalized odds, disparate impact ratio (Phase 2)
Accountable	The system has clear ownership, audit trails, and governance	Model cards, risk registers, incident response playbooks, audit logs

These are not independent
The 7 characteristics interact. A more explainable model (decision tree) may be less accurate (valid & reliable). A privacy-enhanced model (differential privacy) trades utility for privacy. A fair model (equalized odds) may sacrifice overall accuracy. The Responsible AI Engineer's job is not to maximize all 7 — it's to make the tradeoffs explicit, measure them, and document why the chosen balance is appropriate for the specific use case.

Unlock the full lesson

You've read the first 2 sections. The rest of this lesson covers The Three Key Tradeoffs, Building a Trustworthiness Scoring Rubric, What's Next — plus a hands-on lab, quiz, and project artifact.

Create a free account to unlock Phase 0 and Phase 1 of every course — no credit card.