Transparency makes an AI system’s decisions, processes, and relevant information understandable and inspectable. A warning that AI can make mistakes is not enough: people also need to know which documents the system can access, who may use it, and what constraints apply. section 8.2.1
Make transparency practical by documenting and communicating:
- The model’s approved use cases, development process, data sources, preprocessing, access rights, and output-review procedures. section 8.2.2
- Performance evidence, especially response accuracy and relevance, with regular evaluation to find misleading, incorrect, or biased outputs. section 8.2.2
- The model’s limitations, uncertainties, verification method and frequency, and any human adjustments or overrides. section 8.4.4
A transparent AI result
A stakeholder report should identify which conclusions came directly from the model, explain whether people or automated checks verified them, state how often verification occurs, and disclose any human changes. This context helps decision-makers judge uncertainty instead of treating generated material as established fact. section 8.4.4
Transparency does not require exposing every internal model technique. Clear documentation of functionality and data processing can improve trust, support legal and compliance obligations, and reduce risk even when technical explainability is limited. section 8.7.3
