Search Documentation
Search through documentation, guides, and resources.
AI Detection
Understanding how AI detectors work and what results mean
Last Updated: January 27, 2025
How AI Detectors Work
AI detection tools analyze text for patterns that are characteristic of AI-generated content. They use machine learning models trained on millions of examples to identify subtle linguistic signatures.
Detection Methods
Perplexity Analysis
Measures how predictable the text is. AI-generated text tends to be more predictable, choosing common word sequences over creative variations.
Low Perplexity = More predictable = Likely AI
High Perplexity = More varied = Likely Human
Burstiness Detection
Analyzes sentence length variation. Humans naturally vary sentence length dramatically, while AI tends toward uniform structures.
Low Burstiness = Uniform sentences = AI signal
High Burstiness = Varied sentences = Human signal
Linguistic Pattern Matching
Identifies specific words and phrases that AI models overuse, such as "delve," "crucial," "comprehensive," "leverage," and transitional phrases.
Common AI Words:
delve, crucial, comprehensive, leverage, paradigm, robust, streamline, optimizeSentence-Level Analysis
Advanced detectors analyze each sentence individually and provide probability scores, helping identify which parts of the text are most AI-like.
Truuly's Detection Engines
We use 4 different detection engines to provide comprehensive analysis:
GPTZero
Industry LeaderProvides sentence-level detection with detailed probability breakdowns. Used by educators and institutions worldwide.
Strengths:
- • Sentence-by-sentence analysis
- • High accuracy on academic text
- • Detailed probability scores
ZeroGPT
Fast & AccurateSpecializes in quick detection with high confidence scoring for AI vs human-written content.
Strengths:
- • Fast processing times
- • Clear AI vs Human breakdown
- • Good for general content
Originality.ai
Professional GradePremium detection service used by content marketers and SEO professionals. High accuracy across content types.
Strengths:
- • Excellent for marketing copy
- • Plagiarism detection included
- • Trusted by publishers
Local Heuristic
No API RequiredOur proprietary pattern-based detector that runs locally without external API calls. Instant results.
Strengths:
- • Instant analysis
- • Privacy-focused (no data sent)
- • Detects 150+ AI phrases
Understanding Detection Scores
0-30%Human Content
Very low AI probability. The text shows strong human characteristics like varied sentence structure, natural vocabulary, and unpredictable patterns.
What this means:
Your content is highly likely to pass AI detection checks. It exhibits natural writing patterns and should not raise red flags.
30-65%Mixed Signals
Moderate AI probability. The text has both AI and human characteristics. Some detectors may flag it while others don't.
What this means:
Results may vary across different detectors. Consider using a higher humanization mode (Aggressive or Stealth) for more reliable passing.
65-100%AI-Generated
High AI probability. The text shows strong AI characteristics like uniform sentences, predictable word choices, and common AI phrases.
What this means:
This content will likely be flagged by AI detectors. Use Stealth mode humanization or manually edit to add more human variation.
Detection Best Practices
✓ Test with Multiple Detectors
Don't rely on a single detector. Different tools may give different results. Truuly's multi-detector approach gives you confidence across the board.
✓ Check Before and After
Always run detection on both your original AI text and the humanized version to verify improvement. You should see a significant drop in AI probability.
✓ Understand Score Context
Remember that no detector is 100% accurate. Even human-written text can sometimes score 10-20% AI. Focus on getting below 30% for safe passing.
✗ Don't Over-Test
Running the same text through detectors repeatedly won't change the score. If you're not satisfied, re-humanize with different settings instead.
✗ Don't Ignore High Scores
If multiple detectors give you 60%+ scores, don't submit that content. Use Stealth mode or add manual edits to reduce AI signals.
Interpreting Results
All Detectors Agree (Low Score)
Example:
- • GPTZero: 12%
- • ZeroGPT: 8%
- • Originality.ai: 15%
- • Local: 18%
- Average: 13.25%
Interpretation:
Excellent result! Your humanized text is passing all major detectors with flying colors. Safe to submit for academic, professional, or any high-stakes use.
Mixed Results
Example:
- • GPTZero: 45%
- • ZeroGPT: 22%
- • Originality.ai: 58%
- • Local: 31%
- Average: 39%
Interpretation:
Borderline result. Some detectors are flagging AI patterns while others aren't. Consider re-humanizing with Aggressive or Stealth mode for consistency.
All Detectors Agree (High Score)
Example:
- • GPTZero: 78%
- • ZeroGPT: 85%
- • Originality.ai: 72%
- • Local: 81%
- Average: 79%
Interpretation:
Strong AI signals detected across all engines. Do not submit this content. Use Stealth mode humanization with Formal tone and Advanced vocabulary for maximum transformation.
Troubleshooting Detection Issues
Still getting high scores after humanization?
- Switch to Stealth mode for maximum transformation (80-95% change)
- Increase vocabulary level to Advanced for more sophisticated language
- Try Formal tone for academic content, Casual for blog posts
- Manually remove any remaining AI buzzwords ("delve", "crucial", "comprehensive")
- Add personal anecdotes or specific examples that AI wouldn't know
Scores vary wildly between detectors?
- This is normal - different detectors use different algorithms
- Focus on the average score across all detectors
- If 3 out of 4 show low scores, you're likely safe
- Target the specific detector used by your institution/employer if known
Human-written text getting flagged?
- AI detectors are not perfect - false positives happen
- Very formal or technical writing can trigger detectors
- Uniform sentence structure can look AI-generated
- Add more sentence variety and conversational elements
Detector API errors or timeouts?
- Some detectors have rate limits - wait a few minutes and retry
- Very long texts may timeout - try breaking into smaller chunks
- The Local Heuristic detector always works as a fallback
- Check your internet connection and try again