2nd/10_Wiki/Topics/AI_and_ML/AI for Social Good.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, inferred_by, 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	inferred_by	tech_stack
wiki-2026-0508-ai-for-social-good	AI for Social Good (AI4SG)	10_Wiki/Topics	verified	self	AI4SG AI for Good social impact AI public-interest AI humanitarian AI SDG AI	none	B	0.85	conceptual	ai4good social-impact sdg humanitarian climate-ai public-interest ai-ethics		2026-05-09	pending	Claude Opus 4.7 (manual cleanup 2026-05-09)	language applicable_to process / multidisciplinary Non-profit Research Government Corporate Social Responsibility

AI for Social Good (AI4SG)

📌 한 줄 통찰 (The Karpathy Summary)

AI 의 commercial 외 사용. 매 UN SDG (climate, health, education, equity) 의 AI 응용. 매 vendor 의 lab + non-profit + government 의 partnership. Hype 보다 partnership + data + sustainability 가 중요.

📖 구조화된 지식 (Synthesized Content)

정의 + scope

AI 의 application 의 social benefit 목표:

• 매 UN SDG (Sustainable Development Goals) 의 mapping.
• Non-profit / NGO / government partnership 의 흔함.
• 매 commercial value < social value.

매 typical area:

• Climate & sustainability.
• Healthcare (특히 underserved).
• Education (digital divide).
• Disaster response.
• Conservation.
• Accessibility.
• Agriculture (food security).

UN SDG 의 AI mapping

SDG 3: Health

• Diagnosis: malaria detection (mobile + ML), TB X-ray screening.
• Outbreak prediction: 매 epidemic 의 early signal.
• Drug discovery: 매 rare disease 의 candidate.
• Mental health: chatbot support (Wysa, Woebot).
• 매 example: Google's diabetic retinopathy screening (India, Thailand).

SDG 13: Climate

• Forest monitoring: 매 satellite imagery 의 deforestation detect.
• Energy optimization: grid balance, demand prediction.
• Climate model: 매 weather / temperature.
• Methane leak detect: satellite + ML.
• 매 example: Google's flood forecasting (India, Bangladesh).

SDG 4: Education

• Personalized learning: Khanmigo, Duolingo Max.
• Translation: real-time multi-lingual.
• Literacy: 매 student 의 reading support.
• Access: low-bandwidth countries.
• 매 example: AI tutor 의 1.7B underserved.

SDG 11: Cities / Disaster

• Disaster routing: 매 evacuation optimize.
• Population displacement: satellite + social media.
• Damage assessment: 매 earthquake / flood.
• 매 example: Google Crisis Response.

SDG 14, 15: Biodiversity

• Species identification: iNaturalist (10M user).
• Anti-poaching: 매 patrol route + acoustic detection.
• Coral reef monitoring.
• 매 example: Wildbook (whale shark identification).

SDG 5, 10: Equity

• Bias detect: 매 system 의 audit.
• Voice for marginalized: low-resource language.
• Accessibility: 매 disability (vision, hearing).
• 매 example: Project Euphonia (atypical speech).

매 organization 의 program

• Google AI for Social Good: $25M+ funding.
• Microsoft AI for Earth / Health / Accessibility.
• IBM Sustainability Accelerator.
• Anthropic Claude for Climate / Health / Education.
• OpenAI Nonprofit grants.
• DeepMind AlphaFold (free): protein structure.
• UNICEF MagicBox.
• Partnership on AI.

매 framework / methodology

Theory of Change

1. 매 social problem 의 root cause.
2. 매 intervention (AI 의 specific role).
3. 매 outcome (short / long-term).
4. 매 measurement.
5. 매 stakeholder (beneficiary, partner, funder).

Co-design

• 매 affected community 의 participation.
• 매 design 의 representation.
• 매 deployment 의 local trust.
• 매 outcome 의 feedback.

→ "Nothing about us without us".

Human Rights Impact Assessment (HRIA)

• 매 AI 의 deployment 의 human rights effect.
• Privacy, freedom of expression, equality.
• UN B-Tech Project.

매 challenge

Data scarcity

• 매 underserved region 의 data 부족.
• 매 sensitive (health) 의 collection 어려움.
• Synthetic data, transfer learning, federated learning.

Sustainability

• 매 pilot 의 funding 끝 → 매 deployment 의 abandon.
• Local capacity building.
• Open-source.

Bias

• 매 training data 의 Western / urban bias.
• 매 underserved 의 misrepresent.
• Local validation.

Ethics / consent

• 매 vulnerable 의 informed consent.
• 매 data sovereignty (indigenous data).
• 매 deployment 의 community approval.

Verification

• 매 claim 의 evidence.
• "AI4SG washing" (marketing 의 hype + reality 부족).
• 매 outcome 의 measurement 어려움.

매 implementation pattern

Phase 1: Discovery

• Problem definition (community + experts).
• Data audit.
• Stakeholder mapping.
• Feasibility.

Phase 2: Co-design

• Local team partnership.
• Iterative prototype.
• 매 community 의 feedback.

Phase 3: Pilot

• Small-scale deploy.
• 매 outcome 의 measurement.
• 매 unintended effect 의 monitor.

Phase 4: Scale

• 매 partner 의 capacity build.
• Open-source 의 enable.
• Sustainability (funding, governance).

Phase 5: Sustain / Transition

• 매 local ownership.
• Continuous improvement.
• 매 exit plan.

Critique

"AI Solutionism"

• 매 social problem 의 root cause 가 social, not technical.
• 매 AI 의 surface fix.
• 매 tech-driven solution 의 limit.

"AI Colonialism"

• 매 Western / Global North 의 deploy + Global South.
• 매 local agency 의 erasure.
• Data extractivism.

"Pilotitis"

• 매 pilot 의 abundance + scale 의 부족.
• 매 academic / company 의 self-promote.
• 매 sustainable impact 의 부족.

→ Critical perspective + design 의 integration 가 답.

💻 패턴 (응용)

Federated learning (privacy)

# 매 hospital 의 own data + central model.
import flwr as fl

class HospitalClient(fl.client.NumPyClient):
    def __init__(self, model, local_data):
        self.model = model
        self.data = local_data
    
    def fit(self, parameters, config):
        self.model.set_weights(parameters)
        self.model.fit(self.data)
        return self.model.get_weights(), len(self.data), {}

# 매 hospital 의 data 가 own.
# 매 model update 의 share.
fl.client.start_numpy_client(server_address='central:8080', client=HospitalClient(...))

→ 매 patient data 의 hospital 의 own. Central model 의 collective learning.

Low-resource translation (NLLB)

from transformers import pipeline

# Meta NLLB 200 language
translator = pipeline('translation', model='facebook/nllb-200-distilled-600M')

# 매 underserved language
result = translator('Hello', src_lang='eng_Latn', tgt_lang='swh_Latn')
print(result)

→ 매 community 의 mother tongue.

Satellite imagery analysis (deforestation)

# 매 region 의 매 month 의 satellite image
# Diff = deforestation rate
import rasterio
from sentinelhub import SHConfig, BBoxSplitter

# Sentinel-2 의 10m resolution
config = SHConfig()
config.sh_client_id = '...'

# 매 area 의 매 month image
images = fetch_sentinel(area, dates=monthly_2024)
deforestation_mask = ml_model.predict(images)

→ Forest watch 의 ML.

Disaster response (population)

# 매 social media + satellite + cell tower data
import pandas as pd

def estimate_displacement(events):
    cell_density_before = load_ctd('before-event')
    cell_density_after = load_ctd('after-event')
    
    # 매 cell 의 population shift
    delta = cell_density_after - cell_density_before
    return delta

→ Refugee / displacement track.

Health (medical imaging, low-resource)

# 매 mobile-friendly model
import tensorflow as tf

model = tf.keras.applications.MobileNetV3Small(weights='imagenet')
# Fine-tune on disease classification

# Quantize for edge
converter = tf.lite.TFLiteConverter.from_keras_model(model)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
quantized = converter.convert()

# 매 doctor 의 phone 의 deploy

→ Off-grid / low-connectivity.

Accessibility (ASR for atypical speech)

# Project Euphonia (Google) 식
# 매 user 의 own data + base ASR
from transformers import WhisperForConditionalGeneration

model = WhisperForConditionalGeneration.from_pretrained('openai/whisper-base')

# Fine-tune on user's own atypical speech
# (small dataset, transfer learning).

→ Cerebral palsy / ALS 의 communication.

Co-design checklist

# Pre-deployment audit
co_design:
  - Local team 의 partnership: Y/N
  - Affected community 의 input: Y/N
  - Pilot 의 small + measurable: Y/N
  - Outcome 의 transparent disclosure: Y/N
  - Local capacity building: Y/N
  - Sustainable funding: Y/N
  - Exit plan / transition: Y/N
  - Open-source / shared: Y/N

Impact measurement

# 매 outcome 의 quantify
class ImpactTracker:
    def __init__(self):
        self.baseline = self.measure_baseline()
    
    def track(self, intervention_period):
        post = self.measure_after()
        delta = post - self.baseline
        
        # 매 confounder 의 control (RCT 가 ideal)
        return {
            'metric': 'lives_saved',
            'baseline': self.baseline,
            'post': post,
            'delta': delta,
            'confidence': self.compute_confidence(),
        }

→ 매 honest reporting (vs hype).

🤔 의사결정 기준 (Decision Criteria)

상황	추천
Problem 가 social structural	AI 의 limit + structural solution
Tech 가 augment	AI4SG 의 perfect fit
Vulnerable population	Co-design + ethics review
매 region 의 data 부족	Federated / synthetic / transfer
Privacy critical	Federated / on-device
Off-grid	Edge / mobile / quantize
Sustainability concern	Local capacity + open-source

기본값: Co-design + impact measurement + sustainability plan + ethics review. 매 pilot 의 scale path.

⚠️ 모순 및 업데이트 (Contradictions & Updates)

• Solutionism vs structural: 매 social problem 의 tech 의 limit.
• Pilot vs scale: 매 academic / company 의 pilot 의 abundance + scale 의 부족.
• Open-source vs sustainability: 매 open 의 funding model 어려움.
• Local vs global: 매 local context 의 specific need vs global model 의 generality.
• Corporate motive: 매 vendor 의 social good 의 marketing vs sincere commitment.
• AI ethics 의 cost: 매 ethics review 의 development friction.
• 매 SDG 의 hype: 매 vendor 의 SDG checkbox + 매 actual impact 의 부족.

🔗 지식 연결 (Graph)

• 부모: AI-Ethics
• 응용: Federated-Learning
• 관련: AI Humanism · AI Accountability · AI 거버넌스 정책(AI Usage Policy)

🤖 LLM 활용 힌트 (How to Use This Knowledge)

언제 이 지식을 쓰는가:

• 매 nonprofit / NGO 의 AI partnership.
• 매 corporate CSR 의 AI program design.
• 매 SDG 의 AI mapping.
• 매 grant proposal 의 framing.
• 매 pilot 의 sustainability planning.

언제 쓰면 안 되는가:

• Specific country 의 regulation (local expert).
• Crisis 의 immediate response (humanitarian agency).
• Technical implementation 의 detail (engineer).
• Cynicism 의 platform (constructive critique 만).

❌ 안티패턴 (Anti-Patterns)

• Solutionism: 매 social problem 의 tech 의 fix.
• Colonial deploy: local agency 의 erasure.
• Pilotitis: 매 pilot 의 scale 의 plan 부족.
• AI4SG washing: marketing 의 hype + reality 부족.
• Co-design 의 token: 매 community input 의 superficial.
• Open-source 의 abandon: maintenance 의 부족.
• Outcome 의 unmeasured: claim 의 evidence X.
• Ethics review 의 skip: vulnerable 의 harm.

🧪 검증 상태 (Validation)

• 정보 상태: verified (concept-level).
• 출처 신뢰도: B (UN Global Pulse, Partnership on AI, Stanford HAI, Google AI for Social Good reports).
• 검토 이유: Manual cleanup. 매 specific 프로그램 의 detail 가 evolving.

🧬 중복 검사 (Duplicate Check)

• 기존 유사 문서: AI Humanism (related), AI-Ethics (parent), AI 거버넌스 정책(AI Usage Policy) (related).
• 처리 방식: KEEP (specific application focus).
• 처리 이유: AI4SG 가 distinct application area + methodology.

🕓 변경 이력 (Changelog)

날짜	변경 내용	처리 방식	신뢰도
2026-05-08	P-Reinforce Phase 1 정규화	UPDATE	A
2026-05-09	Manual cleanup — SDG mapping + code pattern + 비판 + 안티패턴 + co-design 추가	UPDATE	B