Back to Details

ai-writing-detection

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

AI Writing Detection Reference

AI生成文本检测参考手册

Expert-level knowledge base for detecting AI-generated text, compiled from academic research, commercial detection tools, and empirical analysis.
本手册是检测AI生成文本的专业级知识库，整合了学术研究、商用检测工具及实证分析的成果。

Quick Reference: High-Confidence Signals

快速参考：高置信度检测信号

These indicators strongly suggest AI authorship when found together:
当以下指标同时出现时，强烈暗示文本由AI生成：

Vocabulary Red Flags

词汇警示信号

High-signal words (50-700x more common in AI text):
  • "delve", "tapestry", "nuanced", "multifaceted", "underscore"
  • "intricate interplay", "played a crucial role", "complex and multifaceted"
  • "paramount", "pivotal", "meticulous", "holistic", "robust"
  • "stands/serves as", "marking a pivotal moment", "underscores its importance"
Overused phrases:
  • "It's important to note that..."
  • "In today's fast-paced world..."
  • "At its core..."
  • "Without further ado..."
  • "Let me explain..."
See reference/vocabulary-patterns.md for complete lists.
高信号词汇（在AI文本中的出现频率是人类文本的50-700倍）：
  • "delve", "tapestry", "nuanced", "multifaceted", "underscore"
  • "intricate interplay", "played a crucial role", "complex and multifaceted"
  • "paramount", "pivotal", "meticulous", "holistic", "robust"
  • "stands/serves as", "marking a pivotal moment", "underscores its importance"
过度使用的短语：
  • "It's important to note that..."
  • "In today's fast-paced world..."
  • "At its core..."
  • "Without further ado..."
  • "Let me explain..."
完整列表请查看 reference/vocabulary-patterns.md。

Structural Red Flags

结构警示信号

  • Uniform sentence lengths: 12-18 words consistently (low burstiness)
  • Tricolon structures: "research, collaboration, and problem-solving"
  • Em dash overuse: AI uses em dashes in a formulaic way to mimic "punched up" sales writing, especially in parallelisms ("it's not X — it's Y"); swapping punctuation doesn't fix the underlying emphasis pattern
  • Perfect paragraph uniformity: All paragraphs same approximate length
  • Template conclusions: "In summary...", "In conclusion..."
  • Negative parallelisms: "It's not about X; it's about Y"
  • Elegant variation: Cycling through synonyms to avoid repetition
  • False ranges: "From X to Y" with incoherent endpoints
See reference/structural-patterns.md for details.
  • 统一的句子长度：持续保持12-18词（低突发性）
  • 三并列结构：如 "research, collaboration, and problem-solving"
  • 破折号滥用：AI会公式化地使用破折号模仿“强调式”营销写作，尤其在平行结构中（如"it's not X — it's Y"）；替换标点无法改变其底层的强调模式
  • 完美的段落一致性：所有段落长度大致相同
  • 模板化结尾：如 "In summary...", "In conclusion..."
  • 否定式平行结构：如 "It's not about X; it's about Y"
  • 刻意的词汇替换：循环使用同义词避免重复
  • 虚假范围表述："From X to Y" 搭配逻辑不通的端点
详情请查看 reference/structural-patterns.md。

Content Red Flags

内容警示信号

  • Importance puffery: "marking a pivotal moment in history"
  • Ecosystem/conservation claims without citations
  • "Challenges and Future" sections following rigid formula
  • Promotional language: "nestled in", "stunning natural beauty", "boasts"
  • Superficial analyses: "-ing" phrases attributing significance to facts
See reference/content-patterns.md for details.
  • 重要性夸大：如 "marking a pivotal moment in history"
  • 无引用的生态/环保声明
  • 遵循固定模板的“挑战与未来”章节
  • 营销式语言：如 "nestled in", "stunning natural beauty", "boasts"
  • 表层分析：使用“-ing”短语为事实赋予不必要的重要性
详情请查看 reference/content-patterns.md。

Formatting Red Flags

格式警示信号

  • Title Case in all section headings
  • Excessive boldface (every key term bolded)
  • Inline-header lists: 
    **Bold Header**: description
    pattern
  • Emojis in formal content or headings
  • Subject lines in non-email contexts
See reference/formatting-patterns.md for details.
  • 所有章节标题采用标题大小写（Title Case）
  • 过度加粗：每个关键术语都被加粗
  • 内嵌标题列表：如 
    **Bold Header**: description
    格式
  • 正式内容或标题中使用表情符号
  • 非邮件场景中使用主题行格式
详情请查看 reference/formatting-patterns.md。

Markup Red Flags (Definitive)

标记语言警示信号（确定性指标）

  • turn0search0, turn0image0: ChatGPT reference markers
  • contentReference[oaicite:]: ChatGPT reference bugs
  • utm_source=chatgpt.com: URL tracking (definitive)
  • Markdown in wikitext: ## headers, bold, text
  • grok_card XML tags: Grok/X specific
See reference/markup-artifacts.md for details.
  • turn0search0, turn0image0：ChatGPT的参考标记
  • contentReference[oaicite:]：ChatGPT的引用漏洞标记
  • utm_source=chatgpt.com：URL追踪参数（确定性指标）
  • 在维基文本中使用Markdown：如 ## 标题、加粗、text 格式
  • grok_card XML标签：Grok/X专属标记
详情请查看 reference/markup-artifacts.md。

Citation Red Flags

引用警示信号

  • Broken external links that never existed (no archive)
  • Invalid DOIs/ISBNs: Checksum failures
  • Declared but unused references: Cite errors
  • Placeholder values: 
    url=URL
    , 
    date=2025-XX-XX
See reference/citation-patterns.md for details.
  • 不存在的失效外部链接（无存档记录）
  • 无效的DOI/ISBN：校验和验证失败
  • 声明但未使用的参考文献：引用错误
  • 占位符内容：如 
    url=URL
    , 
    date=2025-XX-XX
详情请查看 reference/citation-patterns.md。

Tone Red Flags

语气警示信号

  • Passive and detached voice throughout
  • Absence of first-person pronouns where expected
  • Consistent formality with no stylistic variation
  • Over-politeness and excessive hedging
  • 全程使用被动、疏离的语气
  • 在预期使用第一人称的场景中未出现第一人称代词
  • 语气始终正式，无风格变化
  • 过度礼貌且频繁使用模糊性表述

Detection Methodology

检测方法论

Multi-Layer Analysis Approach

多层分析方法

Layer 1: Technical Artifact Scan (Definitive)
  • Check for turn0search/oaicite markers (ChatGPT)
  • Check for utm_source=chatgpt.com in URLs
  • Check for grok_card tags (Grok)
  • Check for Markdown in non-Markdown contexts
  • If found: Definitive AI involvement
Layer 2: Vocabulary Pattern Matching
  • Scan for overused AI words/phrases
  • Count frequency of flagged terms
  • Look for clusters of high-signal vocabulary
  • Check for importance/symbolism phrases
Layer 3: Structural Analysis
  • Observe sentence length variation (uniform = AI signal)
  • Check paragraph uniformity
  • Identify repetitive syntactic templates (tricolons, negative parallelisms)
  • Look for elegant variation (synonym cycling)
  • Check for false ranges
Layer 4: Content Pattern Analysis
  • Check for importance puffery and promotional language
  • Look for "Challenges and Future" formula
  • Check for ecosystem/conservation claims without citations
  • Identify superficial analyses with "-ing" attributions
Layer 5: Citation Verification
  • Test external links - do they exist?
  • Verify DOI/ISBN checksums
  • Check for declared but unused references
  • Look for placeholder values
Layer 6: Formatting Analysis
  • Check heading capitalization (Title Case = signal)
  • Count bold phrases per paragraph
  • Look for inline-header list patterns
  • Check for emojis in formal content
Layer 7: Stylometric Observation
  • Pronoun usage patterns (missing first-person?)
  • Tone consistency (too uniform = AI signal)
  • Punctuation patterns (em dash overuse? curly quotes?)
Layer 8: Coherence Check
  • Do paragraphs build a coherent argument?
  • Are concepts repeated with different words?
  • Do transitions actually connect ideas?
Layer 9: Confidence Scoring
  • Weight multiple signals together
  • Require corroborating evidence (3+ signals minimum)
  • Apply context-specific adjustments
  • Check for mitigating factors (human signals)
  • Consider ineffective indicators (don't use them)
第一层：技术痕迹扫描（确定性）
  • 检查是否存在turn0search/oaicite标记（ChatGPT专属）
  • 检查URL中是否包含utm_source=chatgpt.com
  • 检查是否存在grok_card标签（Grok专属）
  • 检查非Markdown场景中是否使用Markdown格式
  • 若检测到以上任意项：可确定AI参与生成
第二层：词汇模式匹配
  • 扫描过度使用的AI词汇/短语
  • 统计警示术语的出现频率
  • 寻找高信号词汇的聚集现象
  • 检查是否存在夸大重要性/象征意义的短语
第三层：结构分析
  • 观察句子长度的变化（统一长度=AI信号）
  • 检查段落一致性
  • 识别重复的句法模板（三并列结构、否定式平行结构）
  • 寻找刻意的词汇替换（同义词循环）
  • 检查是否存在虚假范围表述
第四层：内容模式分析
  • 检查是否存在重要性夸大及营销式语言
  • 寻找遵循固定模板的“挑战与未来”章节
  • 检查是否存在无引用的生态/环保声明
  • 识别使用“-ing”短语的表层分析
第五层：引用验证
  • 测试外部链接是否真实存在
  • 验证DOI/ISBN的校验和
  • 检查是否存在声明但未使用的参考文献
  • 寻找占位符内容
第六层：格式分析
  • 检查标题大小写格式（标题大小写=信号）
  • 统计每段中的加粗短语数量
  • 寻找内嵌标题列表模式
  • 检查正式内容中是否使用表情符号
第七层：文体特征观察
  • 代词使用模式（是否缺失第一人称？）
  • 语气一致性（过于统一=AI信号）
  • 标点使用模式（是否滥用破折号？是否使用弯引号？）
第八层：连贯性检查
  • 段落是否构建了连贯的论证？
  • 是否用不同词汇重复同一概念？
  • 过渡句是否真正衔接了观点？
第九层：置信度评分
  • 综合多个信号的权重
  • 要求至少3个相互佐证的信号
  • 结合具体场景进行调整
  • 检查缓解因素（人类写作信号）
  • 排除无效指标（不依赖此类信号）

Model-Specific Patterns

模型专属特征

Different AI models have distinct "fingerprints":
ModelKey TellsTechnical Artifacts
ChatGPT/GPT-4"delve" (pre-2025), "tapestry", tricolons, em dashes, curly quotesturn0search, oaicite, utm_source=chatgpt.com
ClaudeAnalytical structure, extended analogies, cautious qualificationsNone (uses straight quotes, no tracking)
GeminiConversational synthesis, fact-dense paragraphsNone (uses straight quotes, no tracking)
DeepSeekSimilar to ChatGPT, curly quotesCurly quotation marks
GrokX/Twitter integration
<grok_card>
XML tags
PerplexitySource-focused output
[attached_file:1]
, 
[web:1]
tags
Important dates:
  • ChatGPT launched: November 30, 2022 (text before this is almost certainly human)
  • "delve" usage dropped: 2025 (still signals pre-2025 ChatGPT)
See reference/model-fingerprints.md for detailed model patterns.
不同AI模型具有独特的“指纹”：
模型核心识别点技术痕迹
ChatGPT/GPT-4"delve"（2025年前版本）、"tapestry"、三并列结构、破折号、弯引号turn0search, oaicite, utm_source=chatgpt.com
Claude分析性结构、延伸类比、谨慎的限定表述无（使用直引号，无追踪标记）
Gemini对话式整合、事实密集型段落无（使用直引号，无追踪标记）
DeepSeek与ChatGPT类似、弯引号弯引号
GrokX/Twitter集成
<grok_card>
XML标签
Perplexity以来源为核心的输出
[attached_file:1]
, 
[web:1]
标签
重要时间节点：
  • ChatGPT发布时间：2022年11月30日（此时间之前的文本几乎可以确定为人类写作）
  • "delve"使用量下降：2025年（仍可作为2025年前ChatGPT生成文本的信号）
模型专属特征详情请查看 reference/model-fingerprints.md。

False Positive Prevention

误判预防

Critical requirements:
  • Minimum 200 words for reliable analysis
  • Never flag on single indicators alone
  • Use ensemble scoring (multiple signals required)
High false-positive risk groups:
  • Non-native English speakers (61% false positive rate in research)
  • Technical/formal writing
  • Neurodivergent writers
  • Content using grammar correction tools
Ineffective indicators (do NOT rely on these):
  • Perfect grammar alone
  • "Bland" or "robotic" prose
  • "Fancy" or unusual vocabulary
  • Letter-like formatting alone
  • Conjunctions starting sentences
Signs of human writing:
  • Text from before November 30, 2022
  • Ability to explain editorial choices
  • Personal anecdotes with verifiable details
  • Minor errors and natural quirks
See reference/false-positive-prevention.md for detailed guidance.
核心要求：
  • 文本长度至少200字才可进行可靠分析
  • 绝不仅凭单一指标判定
  • 使用综合评分机制（需多个信号佐证）
高误判风险群体：
  • 非英语母语者（研究显示误判率达61%）
  • 技术/正式写作创作者
  • 神经多样性写作者
  • 使用语法修正工具的创作者
无效检测指标（请勿依赖）：
  • 仅凭借完美语法
  • “平淡”或“机械”的文风
  • “华丽”或不常见的词汇
  • 仅凭借类信件格式
  • 以连词开头的句子
人类写作的特征：
  • 2022年11月30日之前的文本
  • 能够解释编辑选择的依据
  • 包含可验证细节的个人轶事
  • 存在微小错误和自然的写作习惯
误判预防详情请查看 reference/false-positive-prevention.md。

Analysis Output Format

分析输出格式

Structure findings as:
**Overall Assessment**: [Likely AI / Possibly AI / Likely Human / Inconclusive]
**Confidence**: [Low / Medium / High]

**Summary**: 2-3 sentence overview

**Evidence Found**:
- [Category]: [Specific indicator] - "[Quote from text]"
- [Category]: [Specific indicator] - "[Quote from text]"

**Mitigating Factors**: [Elements suggesting human authorship]

**Caveats**: [Limitations, alternative explanations]
请按照以下结构整理分析结果：
**整体评估**：[极可能AI生成 / 可能AI生成 / 极可能人类生成 / 无法确定]
**置信度**：[低 / 中 / 高]

**总结**：2-3句话的概述

**检测到的证据**：
- [类别]：[具体指标] - "[文本引用]"
- [类别]：[具体指标] - "[文本引用]"

**缓解因素**：[暗示人类写作的元素]

**注意事项**：[局限性、其他可能的解释]

Key Principles

核心原则

  1. No certainty claims - AI detection is probabilistic
  2. Multiple signals required - Single indicators prove nothing
  3. Context matters - Academic writing differs from blogs
  4. Stakes awareness - False accusations cause real harm
  5. Evolving field - Detection methods require constant updates
  1. 不做确定性断言 - AI检测是概率性的
  2. 需多个信号佐证 - 单一指标无法证明任何结论
  3. 场景至关重要 - 学术写作与博客写作的标准不同
  4. 重视判定影响 - 误判会造成实际伤害
  5. 领域持续演进 - 检测方法需不断更新

Reference Files

参考文件

  • vocabulary-patterns.md - Complete word/phrase lists with frequencies
  • structural-patterns.md - Sentence, paragraph, and discourse patterns
  • content-patterns.md - Importance puffery, promotional language, content tells
  • formatting-patterns.md - Title case, boldface, emojis, visual patterns
  • markup-artifacts.md - Technical artifacts: turn0search, oaicite, Markdown, tracking
  • citation-patterns.md - Broken links, invalid identifiers, hallucinated references
  • model-fingerprints.md - GPT, Claude, Gemini, Grok, Perplexity specific tells
  • false-positive-prevention.md - Avoiding false accusations, ineffective indicators
  • vocabulary-patterns.md - 完整的词汇/短语列表及出现频率
  • structural-patterns.md - 句子、段落及语篇模式
  • content-patterns.md - 重要性夸大、营销式语言及内容识别点
  • formatting-patterns.md - 标题大小写、加粗、表情符号及视觉模式
  • markup-artifacts.md - 技术痕迹：turn0search、oaicite、Markdown、追踪参数
  • citation-patterns.md - 失效链接、无效标识符、虚构参考文献
  • model-fingerprints.md - GPT、Claude、Gemini、Grok、Perplexity的专属识别点
  • false-positive-prevention.md - 避免误判、无效指标相关指南

Sources

资料来源

This knowledge base synthesizes research from:
  • Stanford HAI (DetectGPT, bias studies)
  • GPTZero, Originality.ai, Turnitin, Pangram methodologies
  • Academic papers on stylometry and discourse analysis
  • Empirical studies on detection accuracy and limitations
  • Wikipedia:WikiProject AI Cleanup field guide (2025)
  • Community-documented patterns from Wikipedia editing
本知识库整合了以下机构的研究成果：
  • 斯坦福HAI（DetectGPT、偏见研究）
  • GPTZero、Originality.ai、Turnitin、Pangram的检测方法
  • 关于文体学与语篇分析的学术论文
  • 检测准确性与局限性的实证研究
  • 维基百科：WikiProject AI Cleanup 领域指南（2025）
  • 维基百科编辑社区记录的检测模式