When I'm deciding whether to depend on a project, I'd like to know how it was built. Did a person write this? Did AI write it, with someone reading every change? Or did it go from a prompt to the main branch without anyone looking?
Today you have to guess from the commit history. So I made AI Code Rating, a three-character rating that a project publishes in an ACR.md file at the root of its repository.
| Position | Range | What it says |
|---|---|---|
| Maintainer Expertise | A –E |
How well the person responsible for the code knows this language and domain (in a solo project that's you; on a team, whoever approves changes), from Expert to Non-programmer |
| AI Share | 0 –4 |
How much of the code AI wrote, from none to 1–25%, 26–50%, 51–75%, and 76–100% |
| Oversight | a –e |
How carefully AI changes were checked, from Verified (read line by line and tested) to Unchecked (merged as generated) |
Here are some examples.
A2b`` B0a``0, Oversight is always a. D4e
Every rating has its own page that explains it, such as aicoderating.com/A2b/. The examples page works through common setups, with the reasoning for each position.
AI Share describes how the code was made, not how good it is. An A3a project, where an expert wrote most of the code with AI and checked every change, can be stronger than a C1e project with a little AI code that nobody checked. The rating gives readers the facts and lets them decide what matters to them.
Most AI disclosure standards use a single level, such as "none", "assisted", or "generated". That folds several questions into one answer. The question I think gets missed is who's doing the checking. The same AI-written code means something different when someone who could have written it reviews it than when nobody on the project can read it.
The level is self-assessed against two practical questions. Could the person responsible for the code have written it without AI, and can they explain every line? The Related Work page compares AI Code Rating with AI-DECLARATION.md and other standards.
---
rating: B4b
spec: "0.1"
updated: 2026-11-07
---
**[ACR B4b](https://aicoderating.com/B4b/)** (Maintainer: Experienced · AI Share: 76–100% · Oversight: Reviewed)
This project is rated with [AI Code Rating](https://aicoderating.com/spec/0.1/), spec version 0.1.
## How AI Was Used
A few sentences in your own words.
That's the actual rating of the AI Code Rating site. Claude Code wrote nearly all of its code and drafted most of its text, and I directed every change and read all of it before it went in. My older project DomainMOD, written almost entirely by hand since 2010, is B1b.
ACR.md file, with the rating linked to its page, and a README badge (Markdown, HTML, or reStructuredText) for you.
The badge comes from shields.io, so nothing about your project depends on the AI Code Rating site.
Yes, ratings are self-reported. But an inflated rating is easy to contradict, since commits, pull requests, and issues show how a project really works, and a rating that doesn't match them costs trust. An honest D4c tells readers more than a doubtful B4b.
Once your ACR.md passes the validator, ask to be added to the Rated Projects list on the home page. It's a short form that asks for the project's name, repository, rating, and a one-sentence description.
The spec is version 0.1, under CC BY 4.0, and the site code is MIT. Issues and suggestions are welcome on GitHub.