You begin by identifying all sanctioned and unsanctioned AI models across your organization.
For each AI model identified, you need to document the following details:
- Model Purpose: What is the intended use of the AI model? For example, customer behavior prediction, fraud detection, or content recommendation.
- Training Data: What data was used to train the model? Include sources, volume, and nature of the data.
- Model Architecture: Describe the architecture of the AI model, including any unique features or design considerations.
- Inputs and Outputs: Document what inputs the model requires and what outputs it produces. This can include data formats, types, and expected value ranges.
- Integration Points: Identify how the AI model interacts with other systems within the organization. This includes data feeds, APIs, and any manual processes involved.
Model cards are concise documents accompanying AI models. They detail the model’s intended use, performance evaluation, and other key information. These cards serve various stakeholders:
- AI practitioners gain insights into model applications.
- Developers use them to compare models and guide future iterations.
- Policymakers assess the model’s impact on policy subjects.
- Organizations consider them for tool adoption decisions.
- Privacy professionals evaluate personal data use and consumer effects.
