Us vs. Them Tracks Line-Level Text Provenance Under AI Editing

Introduction

As AI agents become more involved in coding and writing, a subtle but important question becomes harder to ignore: who actually wrote this part of the file? A Git repository records revisions and authors, but it does not directly tell a later tool whether the current opening paragraph, function, or block of text is mainly human-crafted or mostly generated by an agent. Us vs. Them proposes a practical answer by deriving line-level provenance from version history.

The problem it addresses

The project starts from a simple premise: text that a human has written or deliberately edited should be treated with more caution by an agent. It may represent intent, ownership, or design judgment. Agent-produced “slop,” by contrast, is much more available for rewriting, pruning, or replacement.

This distinction matters in agentic development. A developer may have a mostly AI-generated application but still want to establish certain areas of the code as personally owned or carefully designed. In the next agent session, those areas should not be bulldozed without a strong reason. The same applies to documentation: a README might have been generated initially, while the opening paragraphs were later rewritten by a human. An agent can still improve lower sections, but it should think twice before changing the human-shaped opener.

Key ideas

• No special markup required: The tool is designed for ordinary plain text, including Markdown and source files, without inserting provenance annotations into the document itself.
• Authorship comes from history: Each version is assumed to have identifiable authorship, such as a human author or an agent identity.
• Range-based output: Instead of only labeling isolated lines, the tool emits ranges, or “islands,” of human authorship within a broader “sea” of machine-generated text.
• Diff-based behavior: The implementation relies on diffing across versions, while trying to account for joining, splitting, and diluted authorship rather than collapsing everything into one extreme label.
• CLI or library usage: In a Git repository, users can specify which authors count as “ours” or “theirs”; the shorter side can be named, and the other side is inferred.

Its sample output uses scores such as 1.00 for fully human-authored ranges, 0.00 for fully agent-authored ranges, and intermediate values for text that began as human-authored but was modified by agents to some degree.

Why it matters

Us vs. Them is less about proving an absolute truth and more about giving AI tools a sense of editorial boundaries. If an agent can see that a section reflects human authorship, it can propose changes more carefully, ask for confirmation, or preserve that area unless there is a compelling reason to modify it.

That points toward a broader shift in developer tooling. Future AI coding assistants may need to understand not only syntax, tests, and architecture, but also collaboration history. Provenance, ownership, and editing etiquette could become essential features of agentic IDEs, documentation systems, and automated refactoring workflows.

Source: Hacker News