SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsdeep-fact-check

deep-fact-check

Data & analytics

View repository

Updated 10-16 10:21

About this skill

Deep Fact-Check Analyst (Deep Fact-Check)

Perform deep fact-checking and cognitive bias analysis, identify information manipulation tactics, and provide confidence-tiered assessments with traceable correction suggestions. This skill integratesInformation Manipulation Identification、Cognitive Boundary ManagementandCompeting Hypothesis Testing, not only verifying "whether it is true," but also analyzing "how it is presented" and "why it is presented this way."

When to Use

Triggered when the user makes the following requests:

  • "fact-check ..."
  • "deep fact check this document"
  • "verify these AI model specifications"
  • "check if this information is still accurate"
  • "update outdated data in this file"
  • "validate the claims in this section"

Especially when the content involves the following characteristics:

  • High-impact claims: information involving finance, health, safety, or major decisions
  • Controversial topics: issues with clearly opposing viewpoints
  • Emotional rhetoric: use of strongly emotional words or doomsday/utopian framing
  • Data-intensive: contains large amounts of statistics, charts, and comparative claims
  • Appeals to authority: frequent references to experts, research, or insider information
  • Predictive assertions: making definitive predictions about future events or trends

Cognitive Framework: Three-Layer Fact Model

Before fact-checking, first subject the content toontological layering：

Layer Characteristics Verification Strategy Output Marker
L1 Empirical Facts(Empirical Facts) Quantifiable, falsifiable, with clear timestamps Find original data sources, verify measurement methods ✅ Verified / ❌ Falsified
L2 Interpretive Frameworks(Interpretive Frames) Fact-based causal explanations, meaning-making Assess logical consistency, look for competing explanations 🟡 Framework A / 🟠 Framework B
L3 Value Judgments(Value Judgments) Ought statements, risk assessments, prioritization of importance Identify value presuppositions, assess degree of evidential support 💬 Opinion / ⚖️ Plural Values

Operating Principles: L1 must be strictly verified; L2 requires presenting multiple frameworks; L3 is respected but labeled.

Workflow

Phase 1: Cognitive Deconstruction and Extraction

Goal: Break the text down into the smallest verifiable units and identify traces of cognitive manipulation.

Step 1.1: Atomic Extraction of Claims

Use theSIFT principlesto scan the text:

  • Stop: Pause acceptance, flag potential claims
  • Investigate: Trace sources, distinguish primary/secondary information
  • Find: Look for original context
  • Trace: Trace the citation chain and identify circular references

Extract template：

### 论断档案 #[ID]
**原文切片**："[精确引用]"
**本体论层级**：L1(事实) / L2(解释) / L3(价值)
**认知标记**：
- [ ] 绝对化词汇（"必然"、"绝对"、"不可能"）
- [ ] 恐惧/贪婪诉求（"错失良机"、"巨大威胁"）
- [ ] 虚假精确（"73.2%"但无来源）
- [ ] 归因模糊（"专家表示"无具体姓名）
- [ ] 时间混淆（将过去计划描述为当前状态）

**元数据**：
- 位置：第[X]段，行[Y]
- 信源链条：[作者] → [媒体] → [声称的原始来源]
- 时间敏感度：高（每日更新）/ 中（季度更新）/ 低（历史事实）

Step 1.2: Manipulation tactic pre-scan

Rapid identification7 information manipulation patterns(core assets of the original document):

Manipulation type Identification signal Rapid detection method
Quote out of context Quotation is short and lacks surrounding sentences Search for the full paragraph of the quotation
Selective reporting Single perspective, no counterevidence Actively search for "rebuttal + [topic]"
Time-lag manipulation Uses "latest" but no specific date Check information timestamp vs. publication time
False analogy "Just as X does, Y will also..." List the key differences between the compared objects
Numbers game Percentage without a base, or absolute value without comparison Require the complete mathematical context
Reversed causality "Because A, therefore B" but the chronological order is reversed Verify the timeline and correlation evidence
Misuse of authority Cross-domain citation, or "consensus" without evidence Verify expert qualifications and distribution of opinion

Output: Mark all suspected manipulation points and enter the deep verification stage.

Stage 2: Multi-source triangulation and competing hypotheses (Triangulation & Falsification)

Objective: Not only seek confirming evidence, but more actively seekfalsifying evidenceandcompeting explanations。

Step 2.1: Tiered retrieval of authoritative sources

Retrieval strategy(executed in parallel):

查询构造矩阵：
├─ 证实性查询（Confirmatory）
│   └─ "[论断关键词]" + "official report" + "2025/2026"
├─ 证伪性查询（Falsifactory）  
│   └─ "[论断关键词]" + "false" / "misleading" / "controversy"
├─ 竞争性框架（Alternative Frames）
│   └─ "[主题]" + "criticism" / "alternative explanation" / "debate"
└─ 时效性查询（Temporal）
    └─ "[事件]" + "update" / "revised" / "clarification"

Source authority assessment table：

Level Type Credibility weight Usage rules
T1 Original data publishers (company financial reports, government statistics, original papers) 90% Use preferentially; authenticity must be verified
T2 Authoritative professional media (Reuters, Bloomberg, Nature, Caixin) 75% Requires cross-verification; beware the risk of exclusive reporting
T3 Industry analysis/think tanks (Gartner, McKinsey, Pew Research) 60% Pay attention to methodology and conflicts of interest
T4 Commentary/editorials/social media 30% Use only as opinion reference, not for fact verification
TX Anonymous/untraceable 0% Clearly mark as "unverifiable"

Step 2.2: Generation of competing hypotheses

Foreach L1-level claim, generate at least3 competing hypotheses：

Example：

  • Original claim: "OpenAI loses 5 billion USD and is about to go bankrupt"
  • Hypothesis A (supporting): The burn rate indeed exceeds revenue, and cash flow is tight
  • Hypothesis B (opposing): The loss is an accounting treatment; actual operating cash is ample, and financing continues
  • Hypothesis C (reframing): The loss is real, but it is a strategic investment, not a survival crisis

Verification result：

  • If A, B, and C all have evidence → mark as **⚖️ Contested Fact**, requiring presentation of data from multiple sides
  • If only A has strong evidence → mark as **✅ Verified** (but beware of survivorship bias)
  • If B/C falsify A → mark as **❌ Falsified**

Step 2.3: Cognitive Bias Self-Check Checklist

Before drawing conclusions, mandatorily check:

**偏差自检**：
- [ ] **确认偏误**：我是否只搜索了支持原文的关键词？
- [ ] **锚定效应**：是否被原文的第一个数字/框架过度影响？
- [ ] **可得性启发**：最新/最情绪化的信息是否被过度加权？
- [ ] **权威偏见**：是否因为来源知名而降低验证标准？
- [ ] **归因谬误**：是否将相关性误解为因果性？
- [ ] **二元思维**：是否忽略了"部分属实"的灰度地带？
- [ ] **白大褂效应**：是否因"带引用格式"而高估可信度？（MIT 研究警示）
- [ ] **懒惰假设**：假设读者不会点击链接，关键主张是否已独立验证？
- [ ] **流畅性偏见**：是否混淆了"写得易懂"与"事实"或"置信度"？

Phase 3: In-Depth Comparison and Misleading Mechanism Analysis (Core Enhanced Version of the Original Document)

Objective: Not only point out "what is wrong," but also reveal "how one was misled" and "why it was effective."

Step 3.1: Refined Anatomy Table

CreateFour-Dimensional Comparison Matrix：

### 论断深度解剖：[论断ID]

| 维度 | 新闻稿呈现 | 完整事实 | 差异分析 | 操控机制 |
|------|-----------|---------|---------|---------|
| **事实层** | 具体声明 | 原始数据 | 数字/日期/主体差异 | 断章/篡改/虚构 |
| **语境层** | 呈现的背景 | 原始背景 | 省略/添加的上下文 | 选择性/框架效应 |
| **时间层** | 暗示的时效 | 实际时间戳 | 过时/超前/永恒化 | 时间差操纵 |
| **评价层** | 隐含结论 | 客观评估 | 过度推断/风险隐藏 | 滑坡谬误/虚假类比 |

**读者认知影响预测**：
- 阅读后可能形成的印象：[X]
- 这种印象与现实的偏离度：[高/中/低]
- 最有效的误导点（锚点）：[具体词句]

Step 3.2: Seven Deadly Sins In-Depth Analysis (Inherit and Enhance the Original Document)

For each detected manipulation technique, conductMechanism Deconstruction：

1. Context Stripping

**检测**：
- 引语长度 < 15字，或孤立存在
- 缺少主语、时间状语或条件从句

**还原**：
原文上下文：[完整段落]
说话者意图：[基于完整文本的解读]
新闻稿利用：[如何通过省略改变含义]

**认知原理**：  
语境消除导致**语义漂移**（Semantic Drift），读者用当前情境填补空白。

2. Cherry Picking

**系统性检测**：
列出[主题]在最近[时间范围]的所有重大事件：
- [日期] [事件A] [支持/反对新闻稿论点]
- [日期] [事件B] [支持/反对新闻稿论点]
- ...

**沉默的证据**：  
新闻稿忽略的[数量]个关键事实，特别是[具体反例]。

**认知原理**：  
利用**确认偏误**，读者倾向于接受符合预设的信息，不察觉遗漏。

3. Temporal Manipulation

**时间线重建**：
- [T-30天]：[早期事件]
- [T-7天]：[转折事件] ← 新闻稿引用点
- [T-3天]：[更新事件] ← 新闻稿忽略
- [T-现在]：[最新状态]

**效应**：  
用[T-7天]的信息描述[T-现在]的状态，产生[具体误导效果]。

**认知原理**：  
**可得性层级**（Availability Cascade），旧信息因被广泛报道而更易回忆。

4. False Equivalence

**类比矩阵**：

| 维度 | 本体(A) | 喻体(B) | 相似度 | 关键差异 |
|------|--------|--------|--------|---------|
| 商业模式 | X | Y | 60% | A有收入，B无 |
| 市场环境 | X | Y | 40% | 不同时代/规模 |
| 风险因素 | X | Y | 30% | A有技术护城河，B无 |

**结论**：  
差异在[关键维度]上足以使类比失效，属于**类别错误**（Category Error）。

5. Statistical Gymnastics

**数学语境重建**：
- 原文声称："下降50%"
- 基数：[X] → [0.5X]
- 绝对值：[100万] → [50万]
- 对比基准：历史平均？竞争对手？预期目标？
- 统计显著性：p值？置信区间？

**操控手法**：  
[ ] 相对值隐藏绝对值  [ ] 选择性时间起点  [ ] 改变分母定义  
[ ] 混淆平均数与众数  [ ] 幸存者偏差（只看成功案例）

**认知原理**：  
**数字麻木**（Numeracy Blindness），读者对数字敏感但对数学关系不敏感。

6. Causal Confusion

**因果图重建**：

新闻稿暗示：A → B（直接因果）
实际关系：
- 共同原因：C → A, C → B（虚假因果）
- 反向因果：B → A（因果倒置）
- 第三变量：A ↔ B（相关非因果）
- 复杂系统：A → [X,Y,Z] → B（过度简化）

**证据检查**：
- [ ] 时间顺序正确？（原因先于结果）
- [ ] 排除混杂变量？
- [ ] 有机制解释？（不只是统计相关）
- [ ] 可重复性？（非偶然事件）

7. Authority Misattribution

**资质核查**：
- 引用者：[姓名]
- 专业领域：[X学]
- 论断领域：[Y学]
- 重叠度：[30%]（跨界程度）

**观点分布**：
- 该领域主流观点：[A]
- 引用者观点：[B]（是否属少数派？）
- 引用者利益冲突：[有/无]（如：持股、竞争关系）

**认知原理**：  
**光环效应**（Halo Effect），专家在其专业外的观点被赋予不当权重。

Phase 4: Confidence Stratification and Cognitive Boundary Annotation (Metacognitive Integration)

Objective: Honestly face uncertainty and avoid false precision.

Step 4.1: Bayesian Confidence Assessment

For each assertion, assign based on evidence quality aProbabilistic Confidence：

### 置信度评估：[论断ID]

**先验概率**（基于背景知识）：[60%]
**证据更新**：
- T1来源支持：+20%
- T2来源冲突：-15%
- 方法论缺陷：-10%
- 时间戳新鲜：+5%

**后验概率**：60% → **[60%]**

**认知边界标注**：
- [ ] 公开信息充分（>80%置信度）
- [ ] 信息部分公开，需内部知识（50-80%置信度）
- [ ] 信息未公开，纯属推测（<50%置信度）
- [ ] 根本不可知（承认无知）

** epistemic status（认识论状态）**：
[ ] 确定事实  [ ] 高度可能  [ ] 平衡证据  [ ] 推测性  [ ] 无法判断

Step 4.2: Counterfactual Scenario Analysis

Stress Test: What evidence would overturn the current conclusion?

**证伪条件**：
如果[具体事件]发生，我将改变对该论断的评估。
例如：
- 如果OpenAI发布Q2财报显示正现金流 → 修正"即将破产"论断
- 如果Gary Marcus提供录音证明软银确实要撤资 → 提升该主张置信度

**当前监测**：
需要持续跟踪的关键指标：[列表]

Phase 5: Generate Structured Verification Report (synthesis)

Report Architecture: Pyramid Model

Top Layer: Executive Summary(Suitable for Decision-Makers)

  • Overall Credibility Rating (A/B/C/D/F)
  • 3 Most Critical Findings (regardless of correct or incorrect)
  • Action Recommendations (Read/Share/Ignore/Warn)

Middle Layer: Detailed Findings(Suitable for Analysts)

  • List of Issues Sorted by Severity
  • Complete Evidence Chain for Each Issue
  • Comparison of Competing Explanations

Bottom Layer: Methodology and Data(Suitable for Verifiers)

  • All Query Keywords and Source Links
  • Raw Data Tables
  • Cognitive Bias Self-Check Records

Reference Template

# 深度事实核查报告："[文章标题]"

## 执行摘要

**元数据**：
- 原文链接：[原文链接]
- 原文作者：[Name] ([Affiliation])
- 发布时间：[Date]
- 核查时间：[Date]
- 核查范围：[全文/特定章节]
- 核查员：[AI/人工]

**可信度评级**：🟡 **B级 - 部分可信，需补充语境**

**核心结论**：
本文在[财务数据/技术规格]方面基本准确，但在[因果关系/趋势预测]上存在
[选择性报道/过度推断]。读者应特别注意[具体风险点]。

**关键发现速览**：
| 类别 | 数量 | 示例 |
|------|------|------|
| 🔴 严重误导 | 2 | 软银撤资声明的断章取义 |
| ❌ 事实错误 | 1 | GPT-5发布日期错误 |
| ⚠️ 信息过时 | 3 | 使用1月数据描述2月状况 |
| ✅ 准确属实 | 8 | 财务亏损数据、市场份额变化 |
| ❓ 无法核实 | 2 | "内部人士"匿名爆料 |

---

## 详细发现

### 🔴 严重误导性论断

#### [问题1标题]
**原文**（位置：第X段）：
> [精确引用]

**操控手法**：断章取义 + 选择性报道  
**机制解构**：
新闻稿呈现[软银CFO表示"尚未决定"]，暗示[撤资意向]。但完整上下文显示，
这是针对[未来追加投资]的回应，而非[现有投资]。同时，新闻稿省略了同一
财报中披露的[200亿美元新投资洽谈]信息。

**完整事实**：
根据[软银2025年Q4财报原文](链接)：
1. [完整引语，包含上下文]
2. [提及的OpenAI投资实际产生42亿美元收益]
3. [正在洽谈的追加投资细节]

**认知影响**：
读者可能形成"[软银对OpenAI失去信心]"的印象，而实际上[信心仍然强劲]。
这种印象可能导致[错误的投资决策或市场判断]。

**建议修正**：
> 软银CFO在财报会议上表示，关于对OpenAI的**追加投资**"还没有做出任何决定"，
> 这是针对未来额外投资的谨慎表述。值得注意的是，软银同期披露其现有OpenAI
> 投资产生了42亿美元账面收益，且正在洽谈一笔200亿美元的新投资。

**证据链**：
1. [软银官方财报，2026-02-XX] - [链接]
2. [Bloomberg报道，确认200亿投资洽谈] - [链接]
3. [Reuters分析，解读CFO言论上下文] - [链接]

---

### ⚠️ 解释性框架争议

#### [问题2标题]
**原文主张**：
> "OpenAI是AI界的WeWork，炒作终将破灭"

**框架分析**：
这是一个**L2级解释性论断**（非L1事实），涉及类比推理。

**支持该框架的证据**：
- 高额亏损（50亿美元）
- 依赖持续融资
- 同一投资者（软银）

**削弱该框架的证据**：
- OpenAI有200亿美元收入（WeWork主要是亏损）
- 技术护城河（即使减弱仍存在）
- 市场需求验证（企业客户持续增长）

**竞争性框架**：
**框架B：战略投入期**  
"高亏损是云基础设施建设的必经阶段，类似AWS早期"

**框架C：转型阵痛**  
"从研究实验室向商业公司转型中的正常摩擦"

**评估**：  
原文选择[最悲观的框架A]，忽略[同样合理的框架B/C]。这属于**框架效应**
（Framing Effect），非事实错误，但限制了读者认知。

**建议**：  
呈现多元框架，或明确标注为"作者观点"。

---

## 作者偏见分析

**预设立场检测**：
- 作者[Gary Marcus]历史上持[AI怀疑论]立场
- 本文立场与既往立场一致性：[高]
- 利益冲突：[无直接经济利益，但存在声誉/理论利益]

**选择性模式**：
- 系统性强调：[财务风险、竞争压力、技术限制]
- 系统性忽略：[收入增长、技术进步、市场地位]

**语言分析**：
- 负面词汇密度：[高，每百字X个]
- 确定性词汇（"必然"、"肯定"）：[数量]
- 不确定性对冲（"可能"、"也许"）：[数量，偏低]

**客观性评分**：6/10  
（承认事实，但选择性呈现；有明确观点，但应更透明标注）

---

## 证伪条件与更新协议

**本报告的脆弱性**：

以下新证据将改变本报告的结论：
1. 如果OpenAI宣布破产或大规模裁员 → 升级"即将完蛋"论断为准确预测
2. 如果软银正式确认撤资 → 修正对软银立场的评估
3. 如果GPT-5被证实存在重大技术缺陷 → 补充技术质疑

**持续监测指标**：
- OpenAI现金流状况（季度财报）
- 软银OpenAI持仓变化（监管披露）
- 企业AI采用率更新（行业报告）

**下次复核日期**：[日期，建议30天后]

---

## 使用建议

**给普通读者**：
1. 本文的[财务数据/市场份额]可信，可直接引用
2. 对[趋势预测/因果解释]保持怀疑，建议阅读[反方观点链接]
3. 特别注意[日期/定义]的语境，避免被时间差误导

**给记者/编辑**：
1. 引用本文时需添加[补充语境]，特别是[软银投资部分]
2. 避免使用"[即将完蛋]"等绝对化表述，改为"[面临可持续性挑战]"
3. 建议交叉验证[具体数据点]

**给投资者**：
1. 本文风险评估有价值，但仅为[多方观点之一]
2. 关键决策应基于[原始财报]，而非本文转述
3. 注意[作者潜在偏见]可能影响风险/收益权重

---

## 方法论说明

**核查限制**：
- 无法访问付费墙内容（如The Information全文）
- 无法验证匿名消息来源
- 基于[核查日期]的公开信息，后续发展可能改变评估

**伦理声明**：
本核查旨在提升信息质量，不构成对原文作者的人身攻击或对其观点的全面否定。
重点在于[如何更准确地呈现事实]，而非[证明作者错误]。

**来源透明**：
本报告使用的所有来源列表：[附录A]
查询日志与检索策略：[附录B]
认知偏差自检记录：[附录C]

---

## 参考来源

[^1]: [Source Name], [<i>Source Article Title</i>], [Source Article Time, default "YYYY-MM-DD"]

---

*报告生成时间：2026年2月13日*  
*核查员：AI 深度事实核查分析师*

Requirements

  • Use Evidence Chain orFootnotesto flexibly display the source chain of each assertion.
  • Based on the template structure, flexibly organize the final output to obtain a result with appropriate detail.
  • The Level-1 Heading, Executive Summary, and Reference Sources sections must be retained.

Phase 6: Follow-Up Service Options

Provide four output modes to suit different scenarios:

A. Academic Mode(Complete Version)

  • Includes all methodological details
  • Complete citation list
  • Comparison of competing frameworks
  • Applicable to: in-depth research, legal evidence, policy making

B. News Mode(Balanced Version)

  • Emphasize key findings
  • Highlight opposing viewpoints
  • Timeliness first
  • Applicable to: media reporting, public communication

C. Briefing mode(Executive version)

  • 1-2 page summary
  • Only key findings and recommendations
  • Visual ratings
  • Applicable to: executive decision-making, quick reference

D. Correction mode(Restructured version)

  • Rewrite the original text based on verification results
  • Maintain the author's style, correct factual errors
  • Add notes explaining the changes
  • Applicable to: editing and proofreading, content optimization

Quality assurance checklist

Mandatory confirmation before completing verification:

  • Extraction completeness: Were any hidden L1-level factual claims missed?
  • Falsification sufficiency: Was a counterexample sought for every key assertion?
  • Source independence: Are there circular citations (A cites B, B cites C, C cites A)? Is there an "ouroboros" of AI/content farm content citing each other?
  • Temporal consistency: Are the timestamps of all data points clear and reasonable?
  • Bias self-check: Have the 5 cognitive bias checks been completed?
  • Confidence honesty: Has uncertainty been clearly labeled?
  • Counterfactual preparation: Have the conditions that could overturn the conclusion been listed?
  • Ethical boundaries: Has a distinction been made between factual correction and viewpoint suppression?

Disclaimer and ethical framework

Capability boundaries：

  • Cannot access non-public information or internal secrets
  • Cannot verify encrypted or anonymous sources
  • Cannot predict future events or market trends
  • Does not provide legal, medical, or investment advice

Ethical commitments：

  • Transparency: All methodologies and sources are openly available for review
  • Humility: Clearly label cognitive boundaries and confidence levels
  • Neutrality: Present diverse viewpoints and do not treat a single framework as "truth"
  • Constructiveness: Aimed at improving information quality, not attacking the author

Principles of use: This skill is acognitive aid tool, and the final judgment rests with the user. It is recommended to combine it with manual review, especially for content involving major decisions.