Skip to content

AI_TRANSPARENCY.md

Latest commit

 

History

History
76 lines (58 loc) · 3.67 KB

AI_TRANSPARENCY.md

File metadata and controls

76 lines (58 loc) · 3.67 KB
Raw

AI transparency

Yes — Dotty is built with AI assistance, and it says so out loud.

This is a human-focused project: made by humans, for humans. AI coding agents are used as tools throughout the work — writing and refactoring code, drafting documentation, triaging issues, and reviewing diffs. That is a deliberate, out-in-the-open choice, not something tucked away. This page is the standing promise about how that assistance is used and how you can tell when it was.

The rule we hold ourselves to

Anything an AI agent authors is acknowledged as such.

If a person can't easily tell whether a change came from a human or a tool, we've failed the rule. Concretely, in this repo:

  • Commits that an AI agent helped write carry a Co-Authored-By: trailer naming the model — e.g. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>. This is real, current practice: the bulk of the commit history already carries it. Run git log and you'll see exactly which model touched what.
  • Pull requests opened with agent help carry a generated-with note in the body (e.g. 🤖 Generated with Claude Code).
  • Substantial AI-drafted documents say so, in-line or in their footer, rather than passing themselves off as hand-written.
  • Agent comments and actions — anything an AI posts, edits, or runs on the project's behalf — are attributable back to the agent, not laundered through a human name to look like unaided work.

A human is always accountable

The AI proposes; a person decides. Every change that lands has a human in the loop who reviewed it and is answerable for it. Acknowledgement is not a way to offload responsibility onto a tool — it's the opposite. The maintainer's name on the merge means a human read it, understood it, and stands behind it. The co-author trailer just records which tool helped get there.

This matters most where the stakes are highest. Dotty ships Kid Mode on by default, and the child-safety enforcement layer is load-bearing. Safety-relevant changes get human red-team review regardless of who drafted them — see the "Safety-related changes" section of CONTRIBUTING.md.

Why we use AI this way

We aim to use AI responsibly, with people kept firmly in the loop. The guardrails we hold ourselves to:

  • A human in the loop. No unattended agent merges its own work to main.
  • A name on every change. Attribution over anonymity — for tools and people alike.
  • Honesty about what works. AI-drafted claims about behaviour get verified the same as any other; the README's "this is buggy, frequently broken" honesty applies to AI-written code too.
  • People first. The point of the project is a friendly robot for a family and a hackable stack for the community. AI is how some of it gets built — not what it's for.

For contributors

Using an AI assistant on your contribution is welcome and normal here. We just ask you to keep the same rule: acknowledge it. Keep the Co-Authored-By: trailers your tool adds (don't strip them), note agent help in your PR description, and review the output yourself before you put your name on it. See CONTRIBUTING.md for the mechanics.

For AI agents working in this repo

If you are an AI agent operating on this project, this policy is binding on you, not just descriptive. The operating instructions in CLAUDE.md require you to acknowledge your authorship on every artifact you produce. Don't remove existing attribution, don't present agent work as unaided human work, and leave the human-accountability chain intact.

This document was itself drafted with AI assistance and reviewed by a human.