I recently listened to an on-demand webinar recording on “Fostering Trustworthy AI Using Model Cards” hosted by SAS experts Allie DeLonay, Senior Data Scientist with SAS Data Ethics Practice (DEP) and Vrushali Sawant, who is also a data scientist with SAS’s DEP. It was an ask-the-expert webinar where viewers could raise questions regarding SAS’s AI practices on model cards and ethical data usage.

The webinar began with a very inquisitive question on how organizations are building and managing AI models today and whether those models are open-source, SAS based or created using Python. The hosts also asked about the type of environments those models were built in.

Another question that followed was on the indicators of model health and how important those were to organizations, whether organizations preferred accuracy fairness or drift in AI models once those are implemented in production environments. Also, whether organizations evaluated their AI models based on performance metrics which assessed the model’s data management capabilities depending on distribution of variables from data set to the other.

The third question was regarding the type of models organizations are using, whether those are classification, prediction, text or clustering based.

The webinar then proceeds to talk about the need to develop trustworthy AI which has become evident in recent years, and this is mainly attributed to the fact that there are new regulations being made like the European Union Artificial Intelligence (EU AI) Act. (This is the first binding worldwide regulation on AI meant to set a common framework while delivering AI systems). As AI models rapidly get adopted in organizations, they bring various challenges and to counter those, governments around the world have started enacting these rules.

A model card is defined as a documentation tool used to provide essential information on AI models and is created to improve transparency, accountability and understanding of how a model would work. It typically includes model details, intended use, limitations & risks, domain agnostic information and usage & safety guidelines.

It is rather a difficult task for organizations to explain the usage of AI models to stakeholders and therefore appropriate model cards are being developed by software companies to help them understand the precautionary measures needed to be taken while implementing AI models. In this regard, SAS model cards and model managers come into the picture. A model card helps surface the limitations of an AI model and its ethical considerations. Model cards encourage transparency by providing knowledge on the type of data driving the AI, enable explainability by providing the foundation for explaining the model decisions thereby aiding in identifying potential biases that may be present in a model, and provide recommendations on mitigating them.

A SAS model card is analogous to a nutrition label as it conveys all possible information about the AI’s foundation. It consists of 5 sections: Overview, Data Summary, Model Usage, Model Summary, and Model Audit. The overview section summarizes the model’s components from an accuracy perspective, generalizability perspective, and fairness perspective and is well-curated for business analysts & executives. The model usage section helps understand its intended use, expected benefits, out-of-scope use cases, and limitations. The data summary section helps understand about the data the model was trained on providing insights into the decision-making process. The model summary section offers crucial information on model performance. It describes the target variable, algorithm used, accuracy metrics, and information on model interpretability and fairness assessment for sensitive variables. This section is highly curated for technical engineers. The model audit section allows users to monitor the AI model over time ensuring it is accurate, fair, and isn’t subject to model drift.

The SAS model card is specially designed for both technical and non-technical users and is considered to be highly robust, understandable, and easy to use in the industry.

The webinar further dived into the essentials of generating a complete model card which could be a classification or prediction model built using SAS Model Studio or Python. The first step while creating a model card is to register these models into a model manager using appropriate meta-data. Next is to update model properties. Then for the data summary, an analysis is run, and the data is updated in the SAS information catalogue. Users can analyse model performance over time for which data needs to be collected after the model is trained that contains the actual values. Then users need to build KPI rules and run performance monitoring by creating a definition and generating a report. Finally, the model card is reviewable.

One of the hosts then provided a complete tutorial on how a model card is built by demonstrating these steps in detail in the model studio.

The webinar explained the trustworthy AI lifecycle workflow which was developed by SAS experts based on the NIST AI Risk Management Framework. This workflow ensures that users are able to document important decisions while going through the process of building a trustworthy AI model and helps users take those documents as references for the future. In this lifecycle, the following steps are considered:

• Identifying stakeholders
• Documenting projects
• Preparing and assessing data
• Comparing and validating models
• Testing and deploying models
• Monitoring models
• Auditing projects

Finally, the webinar mentioned about the risk management practices needed to address AI systems across the lifecycle. The National Institute of Standards and Technology (NIST) has tailored its AI RMF to include specific goals for risks specific to AI and emphasizes adaptability across AI lifecycle stages such as data processing, training, and real-world applications. The various steps involved are:

• Identifying if the risk is applicable to the AI model generated
• Suggesting suitable actions for mitigating those risks. The users have the option of implementing recommendations or reporting those risks with proper justification
• Users then need to get those actions approved by the model owner for continuing the workflow but if those actions are rejected, the workflow returns back to risk identification.

This ensures that users innovate the AI models responsibly and with precision.

According to Manakh Phukon Analyst at QKS Group, “SAS model cards improve transparency and accountability in AI by documenting key model details like data sources, performance, and ethical considerations, making them accessible for diverse stakeholders. They act as a “nutrition label” for AI, providing insights into the model’s capabilities and limitations, and helping organizations manage AI models responsibly. Aligned with the NIST AI Risk Management Framework, SAS model cards support a comprehensive lifecycle approach, covering data preparation through to deployment and monitoring, helping organizations meet regulatory and ethical standards effectively.”

Author: Manakh Phukon, Analyst, Analytics and Artificial Intelligence