Ethics & Compliance

Applying AI ethics guidelines and responding to AI regulation in Korea and abroad.

Current AI regulatory landscape

    flowchart TD
    A["AI regulatory landscape"] --> B["EU AI Act<br/>Effective 2024"]
    A --> C["Korea's AI Framework Act<br/>In preparation"]
    A --> D["U.S. AI Executive Order<br/>Effective 2023"]
    A --> E["ISO 42001<br/>AI management system"]

    style A fill:#0891B2,stroke:#0E7490,color:#fff
    style B fill:#2563EB,stroke:#1D4ED8,color:#fff
    style C fill:#7C3AED,stroke:#6D28D9,color:#fff
    style D fill:#EA580C,stroke:#C2410C,color:#fff
    style E fill:#16A34A,stroke:#15803D,color:#fff

EU AI Act risk classification

Risk level	Examples	Requirements
Prohibited	Social credit scoring, real-time biometric identification	Banned outright
High-risk	AI for hiring, credit scoring, medical diagnosis	Strict regulation and auditing
Limited-risk	Chatbots, deepfakes	Transparency obligations
Minimal risk	Spam filters, AI in games	Self-regulation

Responsible AI Dimensions

Regulation says what is prohibited; it does not say how to build. The industry has converged on a set of dimensions that turn “responsible AI” into things that can actually be assigned, implemented, and audited. Every one of them is already a control somewhere in this framework — the value of the list is that it shows which of them you have left unattended:

Dimension	The question it asks	Where the control lives
Fairness & bias mitigation	Does the system produce different outcomes for different groups?	Bias review checklist below; drift checks in Monitoring & Observability
Explainability	Can we say why the system produced this output?	AI Auditability decision records; source citation in RAG
Privacy & security	Can the system leak data, or be made to act against its operator?	Guardrails & Security — PII detection and masking, prompt-injection defense
Veracity & robustness	Is the output actually true, and does it hold up under unusual input?	Hallucination detection in Guardrails; AA-Omniscience scoring in AI Model Benchmarking
Transparency	Do people know they are dealing with AI, and on what terms?	Disclosure patterns in AI UI/UX Design; AI Literacy Education
Governance & accountability	Who is answerable when it goes wrong, and how would we know?	The Governance domain as a whole; AI-Native Design Framework for decision records

Two notes on using this as a checklist. The dimensions are not independent — an unexplainable system cannot be audited for fairness, so explainability gates the dimension above it. And each one needs a named owner: a dimension that belongs to everybody is measured by nobody.

Applying AI ethics principles

Anthropic’s Constitutional AI approach

An approach that internalizes ethical principles during model training:

Example principles:
- Do not provide information that is harmful or dangerous to people
- Do not generate discriminatory or biased responses
- State uncertainty explicitly when it exists
- Respect individual privacy

Practical checklist

Bias review:

Audit training data for demographic bias
Test model output fairness across different groups
Monitor for bias on a regular (quarterly) basis

Ensuring transparency:

Disclose to users when content is AI-generated
Provide explanations for AI decisions (explainability)
Disclose data usage and personal-data handling practices

Establishing accountability:

Clearly designate AI system owners and accountable parties
Establish an incident-reporting procedure for AI-related issues
Operate a regular ethics review committee

AI FinOps