references/ai-detection-patterns.md
4.0 KB查看排版效果| 1 | # AI Content Detection Patterns |
| 2 | |
| 3 | Reference for `ai_content_detector.py`. Explains the three detection methods and how to humanize flagged content. |
| 4 | |
| 5 | ## Method 1: Burstiness (sentence length variance) |
| 6 | |
| 7 | **What it measures:** The coefficient of variation (CV) of sentence lengths across the text. |
| 8 | |
| 9 | **Why it works:** Human writers naturally vary between short punchy sentences (4-8 words) and longer explanatory ones (20-35 words). AI tends to produce consistently medium-length sentences (12-20 words), creating a "flat" rhythm. |
| 10 | |
| 11 | | CV Range | Interpretation | AI Probability | |
| 12 | |---|---|---| |
| 13 | | 0.50+ | High variance — natural human rhythm | 0-30% | |
| 14 | | 0.35-0.49 | Moderate variance — could be either | 30-50% | |
| 15 | | 0.20-0.34 | Low variance — suspiciously uniform | 50-80% | |
| 16 | | < 0.20 | Very flat — strong AI signal | 80-100% | |
| 17 | |
| 18 | **How to fix flagged text:** |
| 19 | - Deliberately insert short sentences. "This matters." "Here's why." |
| 20 | - Break one long sentence into two short ones, then follow with a 25+ word sentence |
| 21 | - Use fragments where tone allows. "Not always. But often enough." |
| 22 | - Vary paragraph length too: alternate 1-sentence and 3-4 sentence paragraphs |
| 23 | |
| 24 | ## Method 2: Vocabulary diversity (Type-Token Ratio) |
| 25 | |
| 26 | **What it measures:** The ratio of unique words to total words in sliding 200-word windows. |
| 27 | |
| 28 | **Why it works:** AI models tend to reuse the same "safe" vocabulary — common verbs, generic adjectives, standard connectors. Human writers use more domain-specific terminology, colloquialisms, and varied word choices. |
| 29 | |
| 30 | | TTR Range | Interpretation | AI Probability | |
| 31 | |---|---|---| |
| 32 | | 0.60+ | Rich vocabulary — likely human | 0-25% | |
| 33 | | 0.45-0.59 | Average — could be either | 25-50% | |
| 34 | | 0.35-0.44 | Repetitive — AI-like | 50-75% | |
| 35 | | < 0.35 | Very repetitive — strong AI signal | 75-100% | |
| 36 | |
| 37 | **How to fix flagged text:** |
| 38 | - Replace generic verbs ("use", "make", "get") with specific ones ("wield", "craft", "extract") |
| 39 | - Add domain jargon where your audience expects it |
| 40 | - Vary connectors: instead of always "however", try "still", "yet", "that said", "then again" |
| 41 | - Remove filler phrases that inflate word count without adding meaning |
| 42 | |
| 43 | ## Method 3: Known AI phrases |
| 44 | |
| 45 | 30 phrases that appear disproportionately in LLM-generated text. These aren't wrong individually — some are perfectly fine in context — but high density signals AI origin. |
| 46 | |
| 47 | **Density thresholds:** |
| 48 | | Density (per 1K words) | Interpretation | |
| 49 | |---|---| |
| 50 | | 0-2 | Normal range | |
| 51 | | 3-5 | Elevated — review flagged phrases | |
| 52 | | 6+ | High — likely AI-generated | |
| 53 | |
| 54 | **The 10 most common AI phrases to watch:** |
| 55 | |
| 56 | 1. "In today's digital landscape" — replace with specific context |
| 57 | 2. "It's worth noting that" — just state the fact |
| 58 | 3. "Leverage" — use "use" unless specifically about financial leverage |
| 59 | 4. "Delve into" — use "explore", "examine", or "look at" |
| 60 | 5. "Game-changer" — use a specific description of impact |
| 61 | 6. "Comprehensive guide" — be specific about what's covered |
| 62 | 7. "Seamlessly integrate" — describe the actual integration |
| 63 | 8. "Robust solution" — describe what makes it robust |
| 64 | 9. "Cutting-edge" — name the specific advancement |
| 65 | 10. "Empower you to" — just say what it enables |
| 66 | |
| 67 | **The fix:** Replace generic AI phrases with specific, concrete language. "In today's digital landscape" → "Since Google's March 2025 core update". "Leverage AI tools" → "Use GPT-4 for first-draft outlines". |
| 68 | |
| 69 | ## Composite scoring |
| 70 | |
| 71 | ``` |
| 72 | Composite = Burstiness × 0.35 + Vocabulary × 0.30 + Phrases × 0.35 |
| 73 | |
| 74 | 0-20: LIKELY_HUMAN — no action needed |
| 75 | 21-50: MIXED — review flagged passages, humanize selectively |
| 76 | 51-100: LIKELY_AI — significant rewriting recommended |
| 77 | ``` |
| 78 | |
| 79 | ## Important caveats |
| 80 | |
| 81 | - This is heuristic, not proof. Technical documentation often has low burstiness and TTR naturally. |
| 82 | - Some AI phrases are perfectly appropriate in context. Don't mechanically remove them all. |
| 83 | - The goal is to make content SOUND human, not to prove it IS human. |
| 84 | - Run this tool AFTER writing, not during — it's an editing pass, not a writing constraint. |
文件内容是平台检查时读出的文本，只供查看。平台不会运行包里的任何脚本或程序。
