This project attempts to provide a straightforward manner in which to disclose generative AI usage. We provide six levels of involvement, with corresponding descriptions given below. The system is designed with worst-case in mind, meaning that the level of AI involvement will increase regardless of how often a certain method is used. As an example, say a project has 5000 human-written L0 PRs but one autonomous PR from Claude; this example project will get an L6 ADL.
AI-Badge currently has an L2 ADL.
| Level | Name | Definition |
|---|---|---|
| L0 | Human | No generative AI used at any point in the development process. |
| L1 | Adoptive | No generative AI used, but the underlying packages or tools have some level of AI involvement. |
| L2 | Consultative | Generative AI used for explanations, research, planning, or debugging; no AI-generated code is used. Documentation is written by humans. |
| L3 | Suggestive | Generative AI provides inline completions, localized editing suggestions, or similar contributions within human-directed coding. Any projects where README files and/or the majority of code comments are generated by AI must fall in this category or higher. |
| L4 | Generative | Generative AI creates implementations in response to requests; a human selects and incorporates these outputs. |
| L5 | Agentic | Generative AI executes multiple or multi-step tasks, edits files, or runs tools, but a human controls the overall integration. This level implies a human reviews the vast majority of pushed code. |
| L6 | Autonomous | Generative AI leads the given project, or changes can be integrated/released without human approval or supervision. Any project where less than 2/3 of the source code is directly reviewed by humans or code is pushed directly by agents automatically falls into this category. |
Ideal use cases of this tool include new graduates, students, or anyone else who works on projects where there will be eventual interest in the level of AI/LLM involvement. This is a self-disclosure tool; it does not monitor commits, nor does it provide quantitative descriptors of AI usage.