ZipPy offers a novel, compression-based approach to detecting AI-generated text, aiming for speed and scalability over traditional large-model methods. It's designed for researchers and developers needing a lightweight, embeddable solution for text classification.

How It Works

ZipPy leverages compression algorithms (LZMA, zlib, Brotli) to indirectly measure text perplexity. By "seeding" a compression stream with AI-generated text, it measures how well a sample text compresses against this seeded dictionary. Texts similar in word choice and structure to the AI corpus will compress better, indicating AI generation. This method approximates the high-probability token sequences characteristic of LLM output.

Quick Start & Requirements

• Install: pip3 install thinkst-zippy or via pkgx install zippy.
• Usage: Run via the zippy command-line script, accepting file paths or reading from stdin.
• Dependencies: Python 3.
• Resources: No specific hardware requirements mentioned beyond standard Python execution.
• Links: Blog post, Hack.LU talk

Highlighted Details

• Employs LZMA, zlib, and Brotli compression for detection.
• Offers an "ensemble" mode combining multiple compression engines.
• Provides browser extensions (Chrome, Firefox) for in-browser detection.
• Output is a statistical comparison of compression ratios, not a percentage.

Maintenance & Community

• The project appears to be a research repository with limited information on active maintenance or community channels.

Licensing & Compatibility

• The README does not explicitly state a license.

Limitations & Caveats

ZipPy's effectiveness is limited to English prose and may perform poorly on non-English text, computer code, or highly stylized content like poetry. The confidence scores are raw deltas and not normalized percentages, requiring careful interpretation.