Tool tracks human vs. AI authorship in text using diff-based line-level provenance

Tracking Text Provenance

A new tool has been developed to provide line-level provenance for text, specifically designed to differentiate between human-authored and AI-generated content. This addresses the increasing need to understand the origin of text in environments where AI agents are actively involved in editing and coding.

Why Provenance Matters

With the rise of agentic coding and editing, the question of authorship becomes critical. Text written or edited by a human is often considered more valuable and should be treated with caution by AI agents. Conversely, AI-generated content can be freely modified or replaced. This tool helps establish boundaries, allowing humans to assert ownership over specific parts of a codebase or documentation.

Technical Implementation

The tool operates by analyzing the version history of a text, such as a Git repository, where each revision carries an authorship marker. It uses a diff-based algorithm to identify "islands" of human-authored lines within a "sea" of machine-generated text. The output indicates the degree of human authorship for specific line ranges, from fully human (1.0) to fully agent-authored (0.00), or a mix (e.g., 0.46).

Usage and Application

The tool can be used as a library or a command-line interface (CLI) tool. For CLI use, it requires `bbin` for local installation. Within a Git repository, users can specify human authors using the `--ours` parameter, and the tool will then identify all other authors as agents. This allows for granular tracking of authorship without requiring specific markup in the text itself.