2nd/10_Wiki/Topics/Computer_Science_and_Theory/Automated-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-automated-decision-making	Automated Decision Making	10_Wiki/Topics	verified	self	ADM Algorithmic Decision-Making Automated Decisions	none	A	0.93	applied	decision-systems ml governance fairness eu-ai-act		2026-05-10	pending	language framework Python scikit-learn/PyTorch/Aequitas

Automated Decision Making (ADM)

매 한 줄

"매 algorithm 의 decision authority — 매 human approval 의 X 또는 minimal.". ADM 의 system 의 input 의 받는 → 매 ML / rule / hybrid 의 통한 decision (loan, hiring, sentencing, content mod, dispatch) 의 output. 매 2026 의 EU AI Act (high-risk category) + GDPR Art.22 + Colorado SB205 의 governed.

매 핵심

매 Categories

• Rule-based: explicit if/then (DMN, decision tables).
• ML scoring: classifier / regressor → threshold.
• LLM agentic: tool-use + reasoning loop (Claude / GPT-5).
• Human-in-the-loop (HITL): ML proposes, human approves.
• Human-on-the-loop: ML acts, human monitors / can override.
• Fully automated: no human in critical path.

매 Regulation (2026)

• EU AI Act (entered force 2025): high-risk systems (biometric, hiring, credit, law enforcement) → conformity assessment, transparency, human oversight.
• GDPR Art.22: right to not be subject to solely automated decision with legal effect; right to explanation.
• Colorado AI Act (SB205): 2026 effective — algorithmic discrimination duty for high-risk AI.
• NYC Local Law 144: AEDT bias audit.

매 Components

• Input validation + preprocessing.
• Feature engineering / embedding.
• Model + threshold + calibration.
• Decision policy (action mapping).
• Logging + audit trail.
• Override / appeal channel.

매 응용

1. Credit underwriting (FICO, Upstart).
2. Hiring screening (with caution post-NYC LL 144).
3. Content moderation (Hive, Perspective API).
4. Insurance claims triage.
5. Dispatching (Uber, DoorDash).

💻 패턴

Pattern 1 — Decision policy with audit

import logging, hashlib, json, time

def decide(applicant, model, threshold=0.6):
    score = model.predict_proba([applicant])[0, 1]
    decision = "approve" if score >= threshold else "review"
    audit = {
        "ts": time.time(), "id": applicant["id"],
        "score": float(score), "threshold": threshold,
        "decision": decision, "model_v": model.version,
        "input_hash": hashlib.sha256(json.dumps(applicant, sort_keys=True).encode()).hexdigest(),
    }
    logging.info(json.dumps(audit))
    return decision, audit

Pattern 2 — Bias audit (Aequitas)

from aequitas.group import Group
g = Group()
xtab, _ = g.get_crosstabs(df_with_predictions, attr_cols=["race", "gender"])
print(xtab[["attribute_name", "fpr", "fnr", "tpr"]])

Pattern 3 — Calibration check

from sklearn.calibration import calibration_curve
prob_true, prob_pred = calibration_curve(y_test, y_score, n_bins=10)
# expect diagonal — deviation = miscalibration

Pattern 4 — HITL queue (low-confidence routing)

def route(score, low=0.4, high=0.7):
    if score >= high: return "auto-approve"
    if score <  low:  return "auto-deny"
    return "human-review"

Pattern 5 — EU AI Act conformity log

@dataclass
class ConformityRecord:
    system_id: str
    risk_class: str   # "high" | "limited" | "minimal"
    intended_purpose: str
    training_data_summary: str
    bias_test_results: dict
    human_oversight_design: str
    last_audit: str

매 결정 기준

상황	Approach
High-stakes (credit, hire, health)	HITL + bias audit + appeal
Reversible low-stakes	full auto + sample review
Real-time (dispatch)	full auto + monitoring
Regulated (EU high-risk)	conformity assessment + transparency
Novel domain	shadow mode → HITL → automation

기본값: HITL + audit log + bias monitoring + appeal channel.

🔗 Graph

• 부모: Decision-Theory
• 변형: HITL
• 응용: Content-Moderation
• Adjacent: Algorithmic-Fairness

🤖 LLM 활용

언제: drafting decision policy, reviewing audit logs for anomalies, generating explanation text (Art.22), bias test fixture generation. 언제 X: autonomous high-stakes decision without human in loop, opaque LLM-only path for regulated domain.

❌ 안티패턴

• Black box high-stakes: no explanation = GDPR/AI-Act violation risk.
• No appeal channel: legitimacy collapse.
• Drift unmonitored: production model 의 silent degrade.
• Proxy discrimination: ZIP code → race proxy unlawful.

🧪 검증 / 중복

• Verified (EU AI Act final text, NIST AI RMF, Aequitas docs).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — FULL content (EU AI Act, HITL, bias audit)