Text Detection

Text Detection

Manuscript uses statistical analysis to differentiate human-written text from AI-generated content.

Detection Signals

Sentence Length Variance

Human writing naturally varies in sentence length. AI tends to produce more uniform sentences.

Pattern	Human	AI
Short sentences	Common	Rare
Long sentences	Common	Common
Variation	High	Low

Weight: 0.15

Vocabulary Richness

Humans use personal vocabulary, including rare words, slang, and domain-specific terms. AI uses statistically “safe” common words.

Weight: 0.20

Contraction Usage

Humans naturally use contractions (“don’t”, “I’m”, “we’ll”). AI often uses formal forms (“do not”, “I am”).

Form	Human	AI
”don’t”	Very common	Rare
”do not”	Rare (formal)	Common
”I’m”	Very common	Rare
”I am”	Rare (formal)	Common

Weight: 0.10

AI Phrase Detection

Certain phrases are strong indicators of AI generation:

• “As an AI…”
• “It’s important to note…”
• “I don’t have personal experiences…”
• “Let me break this down…”
• “That’s a great question…”
• “In summary…”
• “Additionally…”

Manuscript maintains a database of 35+ such patterns.

Weight: 0.20

Hedging Language

AI often uses excessive qualifiers and hedging:

• “It’s possible that…”
• “Generally speaking…”
• “In most cases…”
• “It depends on…”

Weight: 0.10

Punctuation Variety

Humans use diverse punctuation (!?;:—…). AI primarily uses periods and commas.

Weight: 0.10

Repetition Patterns

AI tends to repeat structural patterns mechanically. Humans have organic callbacks.

Weight: 0.10

Example Analysis

Human-written text:

I've been thinking about this for days. Can't shake the feeling that
something's off—you know what I mean? The data just... doesn't add up.

Signals detected:

• High sentence variance ✓
• Contractions used (“I’ve”, “Can’t”) ✓
• Diverse punctuation (?, —, …) ✓
• Natural hedging ✓

Verdict: Human (confidence: 0.92)

---

AI-generated text:

It is important to note that this analysis provides valuable insights.
Additionally, the data suggests several key findings. In summary, the
results demonstrate significant patterns that warrant further investigation.

Signals detected:

• Low sentence variance ✓
• No contractions ✓
• AI phrases detected (“It is important to note”, “Additionally”, “In summary”) ✓
• Formulaic structure ✓

Verdict: AI (confidence: 0.94)

API Usage

curl -X POST http://localhost:8080/verify \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Your text content here..."
  }'

Detailed Response

{
  "id": "hm_abc123",
  "verdict": "human",
  "confidence": 0.87,
  "content_type": "text",
  "signals": {
    "sentence_variance": 0.42,
    "vocabulary_richness": 0.78,
    "contraction_ratio": 0.15,
    "ai_phrases_detected": [],
    "punctuation_variety": 0.65,
    "burstiness": 0.38,
    "hedging_score": 0.22
  },
  "processing_time_ms": 8
}

Accuracy Benchmarks

Metric	Value
Accuracy	90.00%
Precision	100.00%
Recall	80.00%
F1 Score	88.89%

Tested on 100 samples (50 human, 50 AI) including GPT-4, Claude, Gemini, and Llama-3 content.

Limitations

• Short texts (<100 words) have lower accuracy
• Heavily edited AI content may evade detection
• Domain-specific jargon can affect vocabulary analysis
• Non-English text support is limited

Best Practices

1. Minimum length: Provide at least 100 words for reliable detection
2. Original content: Detection works best on unedited content
3. Confidence threshold: Consider results <0.7 as uncertain
4. Multiple samples: For important decisions, analyze multiple excerpts