2nd/10_Wiki/Topics/AI_and_ML/Philosophy.md

---
id: wiki-2026-0508-philosophy
title: Philosophy
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Philosophy of AI, Philosophy Basics]
duplicate_of: none
source_trust_level: A
confidence_score: 0.85
verification_status: applied
tags: [philosophy, epistemology, ethics, ai-safety, consciousness]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: ""
  framework: ""
---

# Philosophy

## 매 한 줄
> **"매 reasoning about reality, knowledge, mind, value"**. Philosophy 의 main branches (epistemology, metaphysics, ethics, mind, logic) — 매 modern AI 의 deeply intertwined: 매 knowing claim (epistemology) → ML evaluation, 매 mind claim (consciousness) → AGI/sentience debate, 매 value claim (ethics) → AI alignment / safety.

## 매 핵심

### 매 main branches
- **Epistemology**: knowledge — 매 what can we know, justified true belief.
- **Metaphysics**: existence — 매 what exists, causation, time, identity.
- **Ethics**: value — 매 right/wrong, deontology vs consequentialism vs virtue.
- **Logic**: valid inference — 매 deductive, inductive, abductive.
- **Philosophy of mind**: consciousness — 매 hard problem (Chalmers).
- **Philosophy of science**: 매 falsification (Popper), paradigms (Kuhn).
- **Philosophy of language**: meaning, reference, speech acts.

### 매 epistemology (AI 와 직결)
- Justified True Belief — Gettier 1963 의 challenge.
- Bayesian epistemology — 매 belief 의 probability degree.
- Reliabilism — 매 process 의 reliability 가 중요.
- 매 ML evaluation: 매 model 의 "knowledge" claim 의 epistemic status.

### 매 ethics (AI alignment)
- **Deontology**: rules (Kant). 매 categorical imperative.
- **Consequentialism**: outcomes (Mill, Bentham utilitarianism).
- **Virtue ethics**: character (Aristotle).
- 매 AI safety: 매 RLHF = consequentialist; constitutional AI = deontological hybrid.

### 매 mind & consciousness
- Hard problem (Chalmers 1995): 매 why subjective experience exists.
- Functionalism: 매 mind = function, substrate-independent → AGI sentience plausible.
- Chinese Room (Searle 1980): 매 syntax ≠ semantics.
- Integrated Information Theory (Tononi): 매 consciousness = Φ.
- 매 2026 frontier: 매 LLM consciousness debate (Anthropic Welfare team, Google DeepMind sentience research).

### 매 응용
1. AI ethics frameworks (alignment, fairness).
2. AGI sentience / moral patienthood debate.
3. Epistemic status of model outputs (hallucination as false belief).
4. Decision theory (CDT, EDT, FDT) for AI agents.

## 💻 패턴

### Bayesian belief update
```python
def bayes_update(prior, likelihood_h, likelihood_not_h):
    """P(H|E) = P(E|H)P(H) / [P(E|H)P(H) + P(E|~H)P(~H)]"""
    p_e = likelihood_h * prior + likelihood_not_h * (1 - prior)
    return likelihood_h * prior / p_e

# Belief in hypothesis after evidence
posterior = bayes_update(prior=0.3, likelihood_h=0.9, likelihood_not_h=0.1)
```

### Constitutional AI (deontological rules)
```python
constitution = [
    "Refuse harmful requests.",
    "Do not deceive.",
    "Respect autonomy.",
]
def critique(response, constitution):
    # LLM critiques own response against rules
    return llm.complete(f"Critique: {response}\nRules: {constitution}")
```

### Trolley problem decision theory
```python
# Utilitarian (consequentialist)
def utilitarian(action_outcomes):
    return max(action_outcomes, key=lambda a: sum(a["lives_saved"]))

# Deontological — hard rule against killing
def deontological(actions):
    return [a for a in actions if not a["actively_kills"]]
```

### Epistemic confidence calibration
```python
# Brier score: lower = better calibration
import numpy as np
def brier_score(predicted_probs, outcomes):
    return np.mean((predicted_probs - outcomes) ** 2)
```

### Falsification check (Popperian)
```python
def is_falsifiable(hypothesis):
    """A hypothesis must specify what would refute it."""
    return hypothesis.get("refuting_observation") is not None
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| AI alignment | Hybrid: consequentialist outcome + deontological hard rules |
| Sentience claim | Skeptical default; functionalism if behavior + introspection |
| Truth claim | Bayesian update + Popperian falsifiability |
| Ethical dilemma | Multi-frame analysis (deont + conseq + virtue) |
| Hard problem | Acknowledge open; don't claim solved |

**기본값**: 매 epistemic humility + multi-framework ethics.

## 🔗 Graph
- 부모: [[Knowledge]]
- 변형: [[Epistemology]] · [[Logic]]
- 응용: [[AI-Safety]] · [[AI_Safety_and_Alignment|AI-Alignment]] · [[AI_Safety_and_Alignment|Constitutional-AI]]
- Adjacent: [[Decision-Theory]]

## 🤖 LLM 활용
**언제**: 매 AI ethics framing, alignment design, sentience debate, epistemic claims.
**언제 X**: 매 narrow technical implementation (philosophy 의 abstraction 만 enough).

## ❌ 안티패턴
- **Single-framework dogma**: 매 only utilitarianism → trolley monsters; only deontology → unable to weigh harm.
- **Conflating AI behavior with consciousness**: 매 LLM 의 "I feel" output ≠ proof of feeling.
- **Solving the hard problem casually**: 매 functionalism 의 plausible but not proven.
- **Ignoring philosophy in alignment**: 매 RLHF 의 implicit consequentialism unexamined.
- **Naive realism in ML eval**: 매 benchmark score = "true intelligence" (epistemic naive).

## 🧪 검증 / 중복
- Verified (Stanford Encyclopedia of Philosophy, Russell & Norvig "AIMA" Ch. 27 Ethics).
- 신뢰도 A-.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — branches, AI applications, decision patterns |