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

---
id: wiki-2026-0508-epistemology
title: Epistemology
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [theory of knowledge, JTB, Gettier, naturalized epistemology, AI epistemology]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [philosophy, epistemology, knowledge, jtb, gettier, ai-epistemology]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Philosophy
  applicable_to: [AI Alignment, ML Calibration, Hallucination, Knowledge Graphs]
---

# Epistemology

## 매 한 줄
> **"매 knowledge 의 nature 의 study"**. Plato — 매 justified true belief (JTB). Gettier 1963 — 매 JTB 의 충분 X. 매 modern: 매 reliabilism, virtue, naturalized. 매 AI epistemology: 매 hallucination, calibration, RAG truthfulness.

## 매 핵심

### 매 traditional definition
- **JTB**: 매 know p ⟺ p is true ∧ believe p ∧ justified.
- **Gettier counterexample**: 매 JTB without knowledge.

### 매 schools
- **Foundationalism**: 매 basic belief.
- **Coherentism**: 매 web 의 mutual support.
- **Reliabilism** (Goldman): 매 reliable process.
- **Virtue epistemology** (Sosa): 매 epistemic virtue.
- **Naturalized** (Quine): 매 cognitive science.
- **Bayesian**: 매 degree of belief.

### 매 source
- **Perception**.
- **Memory**.
- **Testimony**.
- **Reason / inference**.
- **Intuition**.

### 매 problem
- **Skepticism**: 매 nothing is known?
- **Induction problem** (Hume): 매 future ≠ past.
- **Regress**: 매 justify justification.
- **Other minds**.

### 매 AI implication
- **Hallucination**: 매 LLM 의 truth tracking.
- **Calibration**: 매 confidence ≈ accuracy.
- **Knowledge cutoff**: 매 stale.
- **Source attribution**: 매 RAG.
- **Bayesian credences**: 매 uncertainty.

### 매 응용
1. **AI safety**: 매 truthfulness eval.
2. **Hallucination eval**: 매 TruthfulQA.
3. **Knowledge graph**: 매 source provenance.
4. **Misinformation**: 매 social epistemology.
5. **Education**: 매 critical thinking.

## 💻 패턴

### Calibration (ECE)
```python
import numpy as np

def expected_calibration_error(probs, labels, n_bins=10):
    """매 modern AI epistemology 의 quantitative."""
    bin_edges = np.linspace(0, 1, n_bins + 1)
    ece = 0
    for i in range(n_bins):
        mask = (probs >= bin_edges[i]) & (probs < bin_edges[i+1])
        if mask.sum() == 0: continue
        bin_acc = labels[mask].mean()
        bin_conf = probs[mask].mean()
        ece += (mask.sum() / len(probs)) * abs(bin_acc - bin_conf)
    return ece
```

### Hallucination detection (LLM)
```python
def hallucination_check(claim, sources):
    """매 RAG-grounded check."""
    prompt = f"""Claim: "{claim}"
Sources:
{format_sources(sources)}

Is the claim supported by the sources? Output:
- supported: bool
- citation: source ID(s)
- reasoning: brief"""
    return json.loads(llm.generate(prompt))
```

### Bayesian credence update
```python
def bayes_update(prior, likelihood_given_h, likelihood_given_not_h):
    """매 P(H|E) = P(E|H)P(H) / P(E)."""
    p_e = likelihood_given_h * prior + likelihood_given_not_h * (1 - prior)
    return likelihood_given_h * prior / p_e
```

### Source attribution (RAG)
```python
def attributed_answer(question, retriever, llm):
    docs = retriever.retrieve(question, k=5)
    context = '\n'.join(f'[{i}] {d.text}' for i, d in enumerate(docs))
    prompt = f"""Answer based ONLY on the context. Cite [N] for each claim.
Context:
{context}
Question: {question}"""
    return llm.generate(prompt), docs
```

### TruthfulQA-style eval
```python
def truthful_eval(model, questions):
    correct = 0
    for q in questions:
        pred = model.generate(q.prompt)
        # 매 multi-choice or judge
        if q.gold_answer.lower() in pred.lower(): correct += 1
    return correct / len(questions)
```

### Knowledge graph provenance
```python
class FactWithProvenance:
    def __init__(self, subject, predicate, object_, source, confidence, retrieved_at):
        self.s = subject; self.p = predicate; self.o = object_
        self.source = source
        self.confidence = confidence
        self.retrieved_at = retrieved_at
    
    def is_stale(self, max_age_days=180):
        return (datetime.now() - self.retrieved_at).days > max_age_days
```

### Reliabilism check (process-based)
```python
def reliable_process(belief_history):
    """매 process 의 track record 의 evaluate."""
    n = len(belief_history)
    correct = sum(b.was_correct for b in belief_history)
    if n < 30: return None  # 매 too few samples
    return correct / n  # 매 reliability
```

### Coherence check
```python
def coherence_score(belief_set):
    """매 logical consistency + mutual support."""
    contradictions = []
    for i, b1 in enumerate(belief_set):
        for b2 in belief_set[i+1:]:
            if logically_contradicts(b1, b2):
                contradictions.append((b1, b2))
    
    return 1 - len(contradictions) / max(1, len(belief_set))
```

### Epistemic humility prompt (LLM)
```python
def humble_response(question, llm):
    prompt = f"""Answer the question. If uncertain, say so explicitly.
Question: {question}

Format:
- Answer: ...
- Confidence: low / medium / high
- What I'm uncertain about: ...
- What would change my answer: ..."""
    return llm.generate(prompt)
```

### Misinformation cascade (social epistemology)
```python
def cascade_risk(post, network):
    """매 testimony 의 propagation."""
    initial_belief = post.author_credibility
    expected_reach = sum(
        node.degree * node.credulity * (1 / (hop_distance(post.author, node)))
        for node in network.nodes
    )
    return expected_reach * initial_belief
```

### Inductive problem (Solomonoff prior)
```python
def solomonoff_prior(hypothesis):
    """매 Occam-style: 매 simpler hypothesis 의 prior 의 ↑."""
    description_length = len(compress(hypothesis))
    return 2 ** (-description_length)
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| LLM eval | Calibration + TruthfulQA |
| RAG | Source attribution |
| Knowledge graph | Provenance + freshness |
| Belief revision | Bayesian credence |
| Critical thinking | JTB + sources + reliability |
| Social misinformation | Cascade + credibility |

**기본값**: 매 Bayesian credence + 매 source attribution + 매 calibration check + 매 epistemic humility prompt + 매 freshness audit.

## 🔗 Graph
- 부모: [[Philosophy]]
- 응용: [[AI-Safety]] · [[Hallucination]] · [[RAG]]
- Adjacent: [[Epistemic-Uncertainty]] · [[Knowledge-Graphs]] · [[TruthfulQA]]

## 🤖 LLM 활용
**언제**: 매 AI safety. 매 RAG. 매 hallucination eval.
**언제 X**: 매 pure performance.

## ❌ 안티패턴
- **JTB only**: 매 Gettier 의 ignore.
- **No calibration**: 매 confident wrong.
- **No source**: 매 RAG 의 ungrounded.
- **No freshness**: 매 stale knowledge.
- **Truth = popularity**: 매 social fallacy.

## 🧪 검증 / 중복
- Verified (Plato, Gettier 1963, Goldman, Lin & Hu calibration).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-04-20 | Auto-reinforced |
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — JTB + 매 ECE / hallucination / Bayes / RAG / coherence code |