Skip to content

09-research-and-citations.md

Latest commit

 

History

History
121 lines (88 loc) · 10.1 KB

09-research-and-citations.md

File metadata and controls

121 lines (88 loc) · 10.1 KB
Raw

09 — Research & Citations

Ghost Protocol is built on reverse-engineered intelligence from 50+ academic papers, commercial tool documentation, and industry analysis. This page provides full citations for every source.

Academic Papers

Detection Methods

  1. Mitchell, E., Lee, Y., Khazatsky, A., Manning, C.D., & Finn, C. (2023). "DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature." Proceedings of the 40th International Conference on Machine Learning (ICML 2023). arXiv:2301.11305

    • Used in: QA Check #4 (Log Probability Curvature)
    • Key finding: AI-generated text occupies negative curvature regions of log probability functions. Perturbation-based comparison detects AI without any training data.

  2. Bao, G., Zhao, Y., Teng, Z., Yang, L., & Zhang, Y. (2023). "Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature." ICLR 2024. arXiv:2310.05130

    • Used in: QA Check #4 (speed optimization of DetectGPT)
    • Key finding: Conditional probability curvature substitutes perturbation with efficient sampling, achieving 75% speedup with better accuracy.

  3. Hans, A., Schwarzschild, A., Cheber, V., et al. (2024). "Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text." ICML 2024. arXiv:2401.12070

    • Used in: QA Check #5 (Cross-Perplexity Ratio)
    • Key finding: Ratio of perplexity to cross-perplexity between two LLMs creates a powerful discriminator. Works on black-box LLMs without model access. Not biased against non-native English speakers.

  4. Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). "GPT Detectors are Biased Against Non-Native English Writers." Patterns, 4(7). arXiv:2304.02819

    • Used in: QA Checks #26-28 (Bias Exploitation)
    • Key finding: AI detectors consistently misclassify non-native English writing as AI-generated due to lower perplexity in simpler vocabulary and syntax. Simple prompting strategies can bypass detectors entirely.

  5. Mustapha, I.K., Osakue, A., & Odiakaose, K. (2024). "StyloAI: Distinguishing AI-Generated Content with Stylometric Analysis." arXiv:2405.10129

    • Used in: QA Checks #13-18 (Stylometric Analysis)
    • Key finding: 31 stylometric features across 6 categories achieve 81-98% detection accuracy using Random Forest classifiers. 12 features are novel to AI detection.

  6. Gehrmann, S., Strobelt, H., & Rush, A.M. (2019). "GLTR: Statistical Detection and Visualization of Generated Text." ACL 2019 System Demonstrations.

    • Used in: QA Check #3 (Token Distribution)
    • Key finding: Color-coding each token by its rank in the model's prediction distribution reveals AI authorship. Most AI tokens cluster in the top-10 predictions.

  7. Sheth, A.P., et al. (2023). "Counter Turing Test (CT2): AI-Generated Text Detection is Not as Easy as You May Think — Introducing AI Detectability Index." EMNLP 2023. arXiv:2310.05030

    • Used in: QA Check #7 (Entropy), general framework
    • Key finding: Introduced burstiness estimation and entropy-based perplexity measures. Demonstrated that detection difficulty varies dramatically across AI models.

  8. Mikros, G. & Koursaris, S. (2023). "AI-Writing Detection Using an Ensemble of Transformers and Stylometric Features." CEUR Workshop Proceedings, Vol. 3496 (IberLEF 2023).

    • Used in: QA Check #10 (Ensemble Learning)
    • Key finding: Combining transformer classifiers with stylometric features via majority voting outperforms either approach alone.

  9. Kumarage, T., et al. (2023). "Stylometric Detection of AI-Generated Text in Twitter Timelines." arXiv:2303.03697

    • Used in: QA Checks #13, #31 (Stylometry, Mixed Authorship)
    • Key finding: Stylometric features are effective for detecting AI-generated tweets and identifying change points where AI generation begins in a timeline.

  10. Paullet, K., et al. (2025). "A Critical Look at the Reliability of AI Detection Tools." Issues in Information Systems, 26(3), 401-412.

    • Used in: General framework, QA Check #35 (Benchmarking)
    • Key finding: No AI detection tool is foolproof. False positives and bias against non-native English writers erode trust.

Watermarking

  1. Aaronson, S. & Kirchner, T. (2023). "SynthID Text: Watermarking and Identifying Text Generated by Large Language Models." Nature. Google DeepMind

    • Used in: QA Check #19 (Watermark Stripping)
    • Key finding: Token-level watermarking via pseudorandom g-functions creates imperceptible but detectable patterns. Robust to mild edits but defeated by thorough rewriting or translation.

  2. Google DeepMind. (2024). "Watermarking AI-generated text and video with SynthID." deepmind.google/blog

    • Used in: QA Check #19
    • Key finding: Over 10 billion pieces of content watermarked with SynthID. Open-sourced via HuggingFace Transformers v4.46.0+.

Code Stylometry

  1. Tossou, A., et al. (2024). "Is This You, LLM? Recognizing AI-written Programs with Multilingual Code Stylometry." hal-04845581
    • Used in: General framework (extends concept to code)
    • Key finding: 84.1% accuracy detecting AI-written code across 10 programming languages using transformer-based encoder classifiers.

Commercial Tool Documentation

Tool URL Used For
GPTZero gptzero.me Perplexity/burstiness methodology, 7-indicator model
Pangram Labs pangram.com Deep learning detection, critique of perplexity methods
Grammarly grammarly.com/ai-detector RAID benchmark #1, pattern matching
Winston AI gowinston.ai Multi-language, OCR, sentence-level analysis
Originality.ai originality.ai Combined AI + plagiarism detection
QuillBot quillbot.com AI-generated vs AI-enhanced distinction
Writer.com writer.com Enterprise multilingual detection
Polygraf.AI polygraf.ai Cognitive Flow Index, token-based analysis
Overchat overchat.ai Model attribution, AI phrase patterns
Google SynthID ai.google.dev/responsible/docs/safeguards/synthid Open-source text watermarking

Bias & Ethics Studies

  • Stanford HAI — "AI-Detectors Biased Against Non-Native English Writers" (May 2023). Link
  • Proofademic — "Understanding False Positives in AI Detection: A Comprehensive Guide" (Dec 2025). Link
  • University of San Diego Legal Research Center — "Problems with AI Detectors: False Positives and False Negatives" (2024). Link
  • Gonzaga University — "AI Detectors: A Guide for Faculty" (2024). Link
  • The Serials Librarian — "AI Detection Unfairly Accuses Scholars of AI Plagiarism" Vol 85, 2025. Link

Industry Benchmarking & Analysis

  • Walter Writes (2026) — "Best AI Detector Tools: Fully Tested & Ranked" (14-tool benchmark). Link
  • Newo.ai (2025) — "Best AI Detectors of 2025" (methodology review). Link
  • Overchat AI Hub (2025) — "We Tested 20+ Tools" (comprehensive analysis). Link
  • Elegant Themes (2025) — "How to Detect AI Writing: Expert Tips". Link
  • Litero Blog (2025) — "How AI Detectors Actually Work: The Truth About Undetectable AI Tools". Link

Technical Methodology Sources

  • Texas Tech University — "Evaluate: Perplexity and Burstiness" (AI Literacy Guide). Link
  • QuillBot Blog — "Burstiness & Perplexity: Definition & Examples" (Aug 2025). Link
  • GPTZero — "What is perplexity & burstiness for AI detection?" (Oct 2025). Link
  • Pangram Labs — "Why Perplexity and Burstiness Fail to Detect AI" (Mar 2025). Link
  • Hastewire — "How AI Detectors Calculate Perplexity and Burstiness". Link
  • Hastewire — "Stylometry in AI Detection: Does It Work?" Link
  • DEV Community — "SynthID Explained: Technical Deep Dive" (Nov 2025). Link
  • DataCamp — "Google's SynthID: A Guide With Examples" (Jul 2025). Link
  • HuggingFace — "Introducing SynthID Text" (Oct 2024). Link
  • IJRASET — "Deep Learning for Stylometry and Authorship Attribution: Literature Review". Link

Last updated: March 2026