AI Transparency
Production-level transparency about how and why AI was used on a project, distinct from tool-level transparency about what data a model was trained on. The paper focuses on a record of AI usage and decision-making on the project.
Definition (from the paper)
- Transparency is usually discussed as a record of what data AI tools were trained on. The paper instead focuses on production-level AI transparency: a record of AI usage and decision-making on the project.
- To be transparent, a record must be created containing accurate information.
Learnings (FRAMES)
- Substance: a simple human-readable record of how AI was used is a useful internal reference; it is hard to judge how “significant” AI usage is without digging into specific inputs/outputs.
- Process: transparency documentation can be started mid-project or even retroactively; decision/tool details can be captured from meeting transcripts and chat logs; information should be reviewed and confirmed by the team before finalisation.
Framework
- The aimici-transparency-framework operationalises transparency via WHAT / WHO / HOW / WHEN / WHERE and three levels of records.
- See also ai-disclosure (the external-facing action transparency supports) and responsible-ai.