The Technical Tapestry: Assessing AI Systems and Components
Check out Responsible Cyber website : Cybersecurity and Risk Management.
Understanding the technical intricacies of AI systems deployed by vendors is essential for a robust assessment. By scrutinizing the underlying technology, we can reveal potential risks associated with AI components in third-party solutions. This article will focus on two critical aspects: dataset attributes and model attributes.
Dataset Attributes
AI systems require a vast amount of data, and it is crucial to have a clear understanding of the attributes of these datasets. When assessing an AI system, your evaluations should provide clarity into the following:
- Data Quality: Assess the quality of the data used in the AI system. Is it accurate, reliable, and up-to-date?
- Training Data Sources: Determine where the training data comes from. Are the sources reputable and diverse?
- Data Ownership: Understand who owns the data used in the AI system. Is it the vendor, a third party, or your organization?
- Data Versioning and Traceability: Determine if the data used in the AI system is versioned and traceable. This ensures accountability and allows for proper auditing.
These dataset attributes are crucial for evaluating the reliability and integrity of the AI system. By understanding these aspects, you can mitigate potential risks and ensure that the data used aligns with your organization’s standards and requirements.
Model Attributes
Once you have transparency into the datasets used in the AI system, it is essential to gain similar clarity into the model itself. Evaluating the model attributes helps uncover critical information about its design and performance. Consider the following:
- Foundational Model: Determine if the model used is a well-established and widely accepted foundational model. This ensures that the model has been thoroughly tested and proven to be effective.
- Learning Method: Understand the learning method employed by the model. Is it supervised learning, unsupervised learning, or a combination of both?
- Biases: Assess potential biases present in the model. AI systems are only as unbiased as the data they are trained on, so it’s crucial to identify and address any biases that may exist.
- Demographic Parity Ratio: Evaluate the demographic parity ratio of the model. This ratio measures whether the model’s predictions are fair and unbiased across different demographic groups.
- Autonomy Level: Determine the autonomy level of the model. Does it require constant human oversight, or can it operate independently?
These model attributes provide valuable insights into the performance, fairness, and reliability of the AI system. Understanding these aspects allows you to make informed decisions about its suitability for your organization’s needs.
This technical tapestry of dataset and model attributes forms the first layer of a comprehensive evaluation of your third-party AI system. Even though you may not be developing or providing the system yourself, as a deployer of a third-party AI system, you still have obligations and responsibilities regarding the data and models you use. Therefore, it is crucial to have the answers to these questions well-documented.