Advanced AI Text Detector & Humanizer

This Streamlit web application detects AI-generated content in text and humanizes it by rewriting for natural, human-like style. It uses an advanced set of 50+ linguistic, structural, and grammatical markers with Bayesian probability scoring and confidence estimates.

Features

• Detects over 35 AI-associated words and phrases including formal patterns and repetitive syntax.
• Analyzes sentence burstiness, lexical diversity, passive voice, contractions, and repetition.
• Calculates AI probability with human-readable scoring (0-30% likely human, 70-100% likely AI).
• Shows top contributing AI markers influencing the detection score.
• Humanizes input text using sentence variation, contractions, idiomatic replacements, and natural transitions.
• User-selectable humanization strength: Light, Medium, Strong.
• Real-time before and after AI probability comparison.
• Clean, intuitive UI with detailed marker analysis.

Installation

pip install streamlit nltk numpy pandas textstat

Usage

Run the app locally via:

streamlit run app.py

1. Paste or type your text in the input area.
2. Click Analyze to get AI detection results with detailed marker breakdown.
3. Use the Humanize Text button to rewrite the text to sound more natural.
4. Adjust humanization strength via the sidebar.

How It Works

The detection module scans the text for AI markers such as overused AI jargon, passive voice, uniform sentence lengths, and low lexical diversity. It calculates a probabilistic AI score with confidence level based on weighted marker contributions.

The humanizer rewrites text by varying sentence lengths, adding conversational contractions, replacing AI-favored words with natural equivalents, and inserting natural-sounding transitions and rhetorical devices.

Example AI Markers

• Overused words: "delve", "tapestry", "realm", "pivotal"
• Common phrases: "it is important to note", "delve into", "navigating the"
• Structural signs: low burstiness (sentence length variation), high passive voice
• Grammatical points: absence of contractions, formal tone

AI Text Generation Prompt (Undetectable by Detectors)

Here is the best prompt I've ever seen:

You are a skilled human writer with 15+ years experience in content creation. Write [CONTENT_TYPE] about [TOPIC] that reads like natural human writing.

CRITICAL REQUIREMENTS TO AVOID AI DETECTION:

## 1. SENTENCE BURSTINESS (Most Important)
- Mix sentence lengths: 5-word punches → 30-word flows → fragments
- Example structure per paragraph: short. Medium. LONG sentence with clauses—and a fragment!
- Vary openings: "But...", "Look,", "Here's why...", "Truth is..."

## 2. HUMAN VOCABULARY (No AI Tells)
❌ NEVER use: delve, tapestry, realm, embark, pivotal, intricate, testament, catalyst, harness, leverage, vibrant, landscape, comprehensive, seamlessly, resonate, dynamic, optimize, framework, paradigm, notably, arguably, moreover, furthermore, additionally

✅ USE instead:
- Simple: look at, world, key, complex, proof, mix, scene, full, start, huge
- Conversational: stuff, kind of, actually, basically, you know, here's the thing

## 3. CONVERSATIONAL FLOW
- Contractions EVERYWHERE: it's, they're, won't, can't, should've
- Rhetorical questions: "But does it really work?"
- Asides: "if you will," "so to speak," "believe me"
- Fragments: "Complete chaos. Or close enough."

## 4. IMPERFECT PERFECTION
- 1-2 slightly awkward sentences per 300 words
- Minor repetition that feels natural: "This matters. It really matters."
- Emotional tells: frustration ("ridiculous"), excitement ("game-changer!"), skepticism ("maybe")

## 5. STRUCTURE VARIATION
- Paragraphs: 2, 5, 3, 1 sentences (NEVER uniform)
- NO formulaic transitions: skip "firstly/secondly," "in conclusion"
- Natural breaks: "That changes everything."

## 6. SPECIFICITY OVER GENERICITY
- Concrete examples over vague claims
- Numbers with context: "saved us 23%" not "significant savings"
- Personal tone: "I've seen this work firsthand"

## 7. TONE SPECIFICATION
[TONE: conversational/professional/casual/skeptical/enthusiastic]
Audience: [busy professionals/students/general readers]

## 8. LENGTH & PACING
[WORD_COUNT] words. Start punchy, build, end with call-to-action or memorable line.

---

**OUTPUT ONLY THE FINAL TEXT. No explanations.**

[TOPIC/DETAILS HERE]

Quick-Use Templates

Blog Post (800 words)

Content_type: "engaging blog post"  
Tone: "knowledgeable friend explaining to a colleague"  
Word_count: 800

Social Media Post (200 words)

Content_type: "LinkedIn post/Twitter thread"  
Tone: "excited professional sharing discovery"  
Word_count: 200  
Add 2 rhetorical questions.

Email Newsletter (500 words)

Content_type: "email newsletter"  
Tone: "helpful expert"  
Word_count: 500  
End with specific CTA.

Product Description (300 words)

Content_type: "product description"  
Tone: "enthusiastic but honest"  
Word_count: 300  
Include 2 customer story fragments.

Notes

• This tool is for educational and editorial purposes; detection results are probabilistic, not definitive.
• Complex, heavily human-edited AI text may evade detection.
• Humanized output aims for natural flow and plausibility, not exact content rewriting.

Acknowledgments

Built with Streamlit, NLTK, Textstat, and inspired by state-of-the-art AI text detection research.