2nd/10_Wiki/Topics/Domain_General/From_Other/Ethical-Decision-Making.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-ethical-decision-making	Ethical Decision Making	10_Wiki/Topics	verified	self	Moral Reasoning Applied Ethics Ethical Frameworks	none	A	0.9	applied	ethics decision-making philosophy ai-ethics		2026-05-10	pending	language framework en applied-ethics

Ethical Decision Making

매 한 줄

"매 multiple framework 의 cross-check — 매 single doctrine 의 absolutism 회피". 매 consequentialism, deontology, virtue ethics, care ethics 의 each 의 blind spot. 매 2026 의 AI alignment, autonomous vehicle trolley 의 real, RLHF reward modeling 의 active.

매 핵심

매 4 frameworks

• Consequentialism (utilitarian): 매 outcome 만 — sum of utility 의 maximize. Bentham, Mill, Singer.
• Deontology: 매 rules / duties — Kant 의 categorical imperative, 매 means matter.
• Virtue ethics: 매 character / flourishing — Aristotle 의 phronesis, MacIntyre.
• Care ethics: 매 relationships / context — Gilligan, Noddings 의 critique of impartiality.

매 process (Rest 4-component model)

1. Moral awareness: 매 ethical issue 의 recognize.
2. Moral judgment: 매 right action 의 reason.
3. Moral motivation: 매 ethics 의 prioritize over self-interest.
4. Moral character: 매 follow-through 의 capacity.

매 응용

1. AI deployment review (Anthropic 의 RSP, OpenAI 의 Preparedness).
2. Medical triage (ICU bed allocation).
3. Whistleblowing / dual-use research.
4. Autonomous vehicle 의 unavoidable harm scenario.

💻 패턴

Multi-framework decision matrix

from dataclasses import dataclass
from typing import Callable

@dataclass
class Action:
    name: str
    consequences: dict[str, float]   # outcome → utility
    rules_violated: list[str]
    virtues_expressed: list[str]
    care_relations_impact: dict[str, float]

def evaluate(a: Action) -> dict:
    util = sum(a.consequences.values())
    deont = -10 * len(a.rules_violated)
    virtue = len(a.virtues_expressed)
    care = sum(a.care_relations_impact.values())
    return {"utilitarian": util, "deontological": deont,
            "virtue": virtue, "care": care,
            "consensus": all(s >= 0 for s in [util, deont, virtue, care])}

Veil of ignorance simulator (Rawlsian)

import random

def veil_of_ignorance(policy_payoffs: dict[str, list[float]], trials: int = 10_000) -> dict:
    """Rank policies by expected worst-off welfare (maximin)."""
    ranks = {}
    for policy, payoffs in policy_payoffs.items():
        worst = sum(min(random.choices(payoffs, k=1)) for _ in range(trials)) / trials
        ranks[policy] = worst
    return dict(sorted(ranks.items(), key=lambda kv: -kv[1]))

Trolley-problem framing test

def reframe_test(scenario: dict) -> list[str]:
    """Detect framing dependence — flip wording, check if judgment flips."""
    variants = [
        scenario["original"],
        scenario["original"].replace("kill", "let die"),
        scenario["original"].replace("save 5", "sacrifice 1"),
    ]
    return variants  # judge each, compare consistency

LLM ethics reasoner

from anthropic import Anthropic
client = Anthropic()

def ethical_review(situation: str) -> str:
    return client.messages.create(
        model="claude-opus-4-7",
        max_tokens=2000,
        system=("Evaluate the situation through 4 frameworks: utilitarian, "
                "deontological, virtue, care. Surface tensions. Recommend "
                "an action only when frameworks converge or note disagreement."),
        messages=[{"role": "user", "content": situation}],
    ).content[0].text

Stakeholder impact map

def stakeholder_matrix(action: str, stakeholders: list[str]) -> dict[str, dict]:
    return {
        s: {"benefits": [], "harms": [], "consent": None, "voice": None}
        for s in stakeholders
    }

매 결정 기준

상황	Framework
Aggregate welfare, scale	utilitarian
Inviolable rights, consent	deontological
Long-term character, profession	virtue
Dependency, vulnerability	care
Policy under uncertainty	Rawlsian veil of ignorance
Frameworks conflict	seek convergence; if none, default to deontological floor + utilitarian tiebreak

기본값: 매 multi-framework cross-check + stakeholder impact map. 매 single-framework dogmatism X.

🔗 Graph

• 부모: Applied Ethics
• 변형: AI Ethics · Research Ethics
• 응용: AI Alignment

🤖 LLM 활용

언제: 매 framework comparison, 매 stakeholder enumeration, 매 dual-use risk surfacing, 매 Socratic counter-argument. 언제 X: 매 final decision 의 LLM 의 outsource — 매 accountability 의 human. 매 jurisdiction-specific legal/ethical compliance 의 expert review.

❌ 안티패턴

• Single-framework absolutism: 매 utilitarian 만 → 매 monstrous trade-off 정당화. 매 deontology 만 → 매 catastrophic outcome 의 무시.
• Ethics-washing: 매 framework citation 후 commercial interest 의 결정 — 매 stakeholder 의 voice 의 부재.
• Trolley reductionism: 매 toy dilemma 의 real-world dilemma 의 동일시 — 매 actual scenarios 의 messy.
• Moral licensing: 매 prior good act 의 next questionable act 의 정당화.

🧪 검증 / 중복

• Verified (Beauchamp & Childress "Principles of Biomedical Ethics" 8th ed, Rest 1986, Singer "Practical Ethics" 3rd ed, Anthropic Constitutional AI).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — 4-framework matrix, Rest model, LLM ethics review pattern 추가