hum-source-criticism

asgard-ai-platform ｜ 除錯與調查 ｜ 原文 ｜ MIT

用外部內部考證和三角驗證判斷資訊來源可不可信

---
name: "hum-source-criticism"
description: "運用一手/二手來源分類、內部/外部考證、三角驗證與假訊息偵測，評估來源的可信度。當使用者需要評估資訊是否可信、評估研究來源、查核主張，或偵測假訊息時使用，即使他們說的是「這個來源可以信嗎」「這是真的嗎」「這份資料有多可靠」或「幫我查核這個」也適用。"
metadata:
  category: "WP-19 文學院/人文"
  tags: ["humanities", "source-criticism", "fact-checking", "information-literacy"]
---

# 史料考證

## 概述

史料考證是一套系統性的方法，用來評估資訊是否可信。它原本來自歷史學方法論，如今在這個充斥假訊息、把意見當事實、以及 AI 生成內容的資訊環境中，已是不可或缺的能力。

## 框架

```
鐵律：沒有任何來源是自動可信的

每個來源，包括學術期刊、政府資料與信譽良好的媒體新聞，
都可能有偏見、錯誤與限制。可信度是被評估出來的，
不是被預設的。「這是《紐約時報》/ 中央社的報導」並不夠，
他們在報導「什麼」？依據「什麼」證據？其他來源有沒有佐證？
```

### 來源分類

**一手來源**：來自當時/事件的直接證據（原始文件、原始資料、目擊者的說法、原創研究、官方紀錄）

**二手來源**：對一手來源的分析或詮釋（教科書、回顧性文章、新聞分析、傳記）

**三手來源**：對一手與二手來源的彙編（百科全書、Wikipedia、資料庫），是起點，不是終點

### 來源可信度的四項檢驗

**1. 外部考證**：來源是真的嗎？
- 誰創作的？他們是他們自稱的那個人嗎？
- 什麼時候創作的？日期前後一致嗎？
- 這是原件，還是被改動過？
- 發表的刊物/平台信譽良好嗎？

**2. 內部考證**：內容可靠嗎？
- 作者在這個主題上有專業嗎？
- 作者可能有什麼偏見或利益？
- 有沒有引用證據？能不能查證？
- 推理合乎邏輯嗎？結論有證據支持嗎？

**3. 三角驗證**：多個獨立的來源是否一致？
- 檢查 3 個以上獨立的來源（不是同一份原始報導的複本）
- 「獨立」是指不同的作者、不同的組織、不同的方法
- 獨立來源之間的一致，會增強信心

**4. 時效性**：資訊夠新嗎？
- 什麼時候發表的？之後情況有沒有改變？
- 對於變化快速的主題（AI、政策、市場），連 6 個月前的來源都可能過時

### 假訊息的警訊

| 警訊 | 說明 |
|----------|-----------|
| 沒有標明作者或組織 | 誰在為這個主張背書？ |
| 有情緒性語言卻沒有證據 | 目的是挑起情緒，不是提供資訊 |
| 沒有引用一手來源 | 主張沒有可追溯的證據 |
| 「研究顯示」卻沒說是哪項研究 | 含糊地訴諸權威 |
| 單一來源在許多網站上被放大 | 同一個主張被複製，而不是獨立查證 |
| 好得不像真的 / 太離譜 | 極端的主張需要極端的證據 |
| 網址/網域模仿有信譽的來源 | Fakecnn.com、bbc-news.co（不是 bbc.co.uk） |

## 輸出格式

```markdown
# 來源評估：{來源/主張}

## 來源身分
- 作者/組織：{誰}
- 刊物：{哪裡}
- 日期：{何時}
- 類型：一手 / 二手 / 三手

## 可信度評估
| 檢驗 | 評估 | 證據 |
|------|-----------|---------|
| 外部（真實嗎？） | ✓/⚠/✗ | {推理} |
| 內部（可靠嗎？） | ✓/⚠/✗ | {推理} |
| 三角驗證（有佐證嗎？） | ✓/⚠/✗ | {查過的其他來源} |
| 時效性（夠新嗎？） | ✓/⚠/✗ | {日期的相關性} |

## 警訊
- {偵測到的任何警訊}

## 結論
- 可信度：高 / 中 / 低
- 建議行動：{信任 / 進一步查證 / 捨棄}
```

## 範例

### 正確的應用
**情境：** 評估一則瘋傳的社群貼文，宣稱「台灣的 GDP 將在 2027 年超越南韓」

| 檢驗 | 評估 | 證據 |
|------|-----------|---------|
| 外部 | ⚠ | 匿名帳號、沒有機構隸屬、圖表沒有資料來源 |
| 內部 | ✗ | 用的是名目 GDP（不是購買力平價）、只挑半導體產業的預測、忽略匯率波動 |
| 三角驗證 | ✗ | IMF 與世界銀行的預測都沒有顯示這樣的趨近，沒有任何有信譽的分析師做出這個主張 |
| 時效性 | ✓ | 本月發布 |

**警訊**：情緒化的標題（「台灣碾壓南韓」）、沒有引用一手資料來源、只有一張沒有出處的圖表
**結論**：可信度低，捨棄 ✓

### 不正確的應用
- 「這是路透社的，所以一定是真的」 → 可信度是被預設的，不是被評估的。就算是有信譽的來源，也可能出錯、過時，或用特定方式框架議題。違反鐵律。

## 常見陷阱

- **有偏見 ≠ 不可靠：** 每個來源都有立場。工會關於工作條件的報告有偏見，但可能包含準確的資料。偏見與準確度要分開評估。
- **Wikipedia 是起點：** 它是附有參考資料的三手來源。順著參考資料找到一手/二手來源。不要把 Wikipedia 當成證據來引用，要引用 Wikipedia 所引用的東西。
- **AI 生成的內容：** AI 可以產生有說服力、卻是捏造的「來源」（假論文、假引述、假統計）。要確認引用的來源真的存在。
- **共識 ≠ 真理，但是強烈的訊號：** 科學共識（氣候變遷、疫苗安全）是目前能取得的最強證據。與共識相左的孤立「專家」，需要非比尋常的證據。
- **來源的可信度取決於領域：** 心臟科醫師在心臟疾病上是可信的來源，在經濟學上則不是。要讓專業與主張相符。

## 參考資料

- CRAAP 測試（Currency、Relevance、Authority、Accuracy、Purpose），見 `references/craap-test.md`
- 事實查核的工具與資料庫，見 `references/fact-check-tools.md`

---
name: "hum-source-criticism"
description: "Evaluate source credibility using primary/secondary classification, internal/external criticism, triangulation, and misinformation detection. Use this skill when the user needs to assess whether information is trustworthy, evaluate research sources, fact-check claims, or detect misinformation — even if they say 'can I trust this source', 'is this real', 'how reliable is this data', or 'fact-check this for me'."
metadata:
  category: "WP-19 文學院/人文"
  tags: ["humanities", "source-criticism", "fact-checking", "information-literacy"]
---

# Source Criticism

## Overview

Source criticism is a systematic method for evaluating whether information is trustworthy. Originally from historical methodology, it's now essential for navigating an information environment flooded with misinformation, opinion-as-fact, and AI-generated content.

## Framework

```
IRON LAW: No Source Is Automatically Trustworthy

Every source — including academic journals, government data, and news from
reputable outlets — has potential biases, errors, and limitations. Credibility
is assessed, not assumed. "It's from the New York Times / 中央社" is not
sufficient — WHAT are they reporting, based on WHAT evidence, and do other
sources corroborate it?
```

### Source Classification

**Primary sources**: Direct evidence from the time/event (original documents, raw data, eyewitness accounts, original research, official records)

**Secondary sources**: Analysis or interpretation of primary sources (textbooks, review articles, news analysis, biographies)

**Tertiary sources**: Compilations of primary and secondary (encyclopedias, Wikipedia, databases) — starting points, not endpoints

### Four Tests of Source Credibility

**1. External Criticism** — Is the source authentic?
- Who created it? Are they who they claim to be?
- When was it created? Is the date consistent?
- Is it the original or has it been altered?
- Is the publication/platform reputable?

**2. Internal Criticism** — Is the content reliable?
- Does the author have expertise in this topic?
- What is the author's potential bias or interest?
- Is the evidence cited? Can it be verified?
- Is the reasoning logical? Are conclusions supported by the evidence?

**3. Triangulation** — Do multiple independent sources agree?
- Check 3+ independent sources (not copies of the same original report)
- "Independent" means different authors, different organizations, different methods
- Agreement across independent sources strengthens confidence

**4. Currency** — Is the information current enough?
- When was it published? Has the situation changed since then?
- For fast-moving topics (AI, policy, markets), even 6-month-old sources may be outdated

### Red Flags for Misinformation

| Red Flag | Description |
|----------|-----------|
| No author or organization identified | Who stands behind this claim? |
| Emotional language without evidence | Designed to provoke, not inform |
| No primary sources cited | Claims without traceable evidence |
| "Studies show" without naming the study | Vague appeals to authority |
| Single source amplified across many sites | Same claim copied, not independently verified |
| Too good to be true / too outrageous | Extreme claims require extreme evidence |
| URL/domain mimics reputable source | Fakecnn.com, bbc-news.co (not bbc.co.uk) |

## Output Format

```markdown
# Source Evaluation: {Source/Claim}

## Source Identity
- Author/Organization: {who}
- Publication: {where}
- Date: {when}
- Type: Primary / Secondary / Tertiary

## Credibility Assessment
| Test | Assessment | Evidence |
|------|-----------|---------|
| External (authentic?) | ✓/⚠/✗ | {reasoning} |
| Internal (reliable?) | ✓/⚠/✗ | {reasoning} |
| Triangulation (corroborated?) | ✓/⚠/✗ | {other sources checked} |
| Currency (current?) | ✓/⚠/✗ | {relevance of date} |

## Red Flags
- {any detected red flags}

## Verdict
- Credibility: High / Moderate / Low
- Recommended action: {trust / verify further / discard}
```

## Examples

### Correct Application
**Scenario:** Evaluating a viral social media post claiming "Taiwan's GDP will surpass South Korea's by 2027"

| Test | Assessment | Evidence |
|------|-----------|---------|
| External | ⚠ | Anonymous account, no institutional affiliation, chart has no data source |
| Internal | ✗ | Uses nominal GDP (not PPP), cherry-picks semiconductor sector projection, ignores exchange rate volatility |
| Triangulation | ✗ | IMF and World Bank projections show no such convergence; no reputable analyst makes this claim |
| Currency | ✓ | Posted this month |

**Red flags**: Emotional headline ("Taiwan DESTROYS Korea"), no primary data source cited, single unsourced chart
**Verdict**: Low credibility — discard ✓

### Incorrect Application
- "This is from Reuters, so it must be true" → Credibility assumed, not assessed. Even reputable sources can be wrong, outdated, or framing an issue in a particular way. Violates Iron Law.

## Gotchas

- **Bias ≠ unreliable**: Every source has a perspective. A labor union's report on working conditions is biased but may contain accurate data. Assess bias AND accuracy separately.
- **Wikipedia is a starting point**: It's a tertiary source with references. Follow the references to primary/secondary sources. Don't cite Wikipedia as evidence — cite what Wikipedia cites.
- **AI-generated content**: AI can produce convincing but fabricated "sources" (fake papers, fake quotes, fake statistics). Verify that cited sources actually exist.
- **Consensus ≠ truth, but it's a strong signal**: Scientific consensus (climate change, vaccine safety) is the strongest available evidence. Lone dissenting "experts" who contradict consensus need extraordinary evidence.
- **Source credibility is domain-specific**: A cardiologist is a credible source on heart disease but not on economics. Match expertise to the claim.

## References

- For CRAAP test (Currency, Relevance, Authority, Accuracy, Purpose), see `references/craap-test.md`
- For fact-checking tools and databases, see `references/fact-check-tools.md`