What's Missing in AI Risk Frameworks? Our Guide to Holistic Assessment

Artificial Intelligence (AI) technologies are revolutionizing various sectors, from healthcare to finance, but they also introduce several risks that cannot be ignored.

A framework that has been commonly referenced for assessing these risks is KAIRI (Key AI Risk Indicators), which maps onto the recently proposed regulatory requirements set forth in the Artificial Intelligence Act [1].

While the KAIRI framework is insightful and aligns with emerging regulations, it may not capture all the nuances and complexities of AI applications across different sectors.

To address this, we propose an extended framework that includes additional dimensions aimed at capturing a more comprehensive picture of the risks associated with AI applications.

The Original KAIRI Framework

• Sustainability: Sustainability is about the robustness and cybersecurity of AI models. Variable selection methods can enhance sustainability by creating parsimonious models that are stable against data variations. Statistical tests like the F-test or the Chi-square test can help measure the significance of the difference between two nested models.

• Accuracy: Accuracy is assessed through predictive validity. Metrics like RMSE (Root Mean Square Error) for continuous variables and AUROC (Area Under the ROC Curve) for categorical variables can be employed, supported by statistical tests like Diebold–Mariano's and DeLong's, respectively.

• Explainability: This component is crucial for all stakeholders. For models that are considered "black-boxes," Shapley values are often employed to gauge the explainability of each variable. Key Risk Indicators based on the percentage of Shapley values can offer deeper insights.

• Fairness: Fairness ensures that AI applications are unbiased across different groups. The Gini coefficient can measure the concentration of a variable's importance among these groups, while the Kolmogorov–Smirnov test can gauge the departure from a uniform distribution, indicating fairness.

Extending the Framework

User Trust

The intent behind adding User Trust is to ensure that the AI system is not only technically sound but also socially and culturally accepted. High-performing algorithms can still fail in the market if users do not trust them.

Why It’s Needed

User trust is crucial for the adoption and long-term success of any AI system. When users trust a system, they are more likely to use it consistently, provide valuable feedback, and even advocate for it. Trust is especially important in sectors like healthcare, finance, and law enforcement, where the stakes are high.

Indicators

Elements such as Social Proof, Certifications by Independent Parties, Historical Performance, Reputation, Third-Party Audits, Trial Periods and Demonstrations, and Error Correction and Accountability serve as robust indicators to measure and build user trust.

Data Privacy and Security

The intent here is to safeguard the sensitive information that AI systems often handle.

Why It’s Needed

With increasing concerns about data breaches and unauthorized access, focusing on cybersecurity and data integrity is non-negotiable. Failure to secure data not only leads to legal consequences but also erodes user trust.

Components

• Cybersecurity: Evaluates the risk of unauthorized access and data breaches.

• Data Integrity: Measures the risk associated with data corruption or loss.

Economic Viability

The aim is to gauge whether the AI system makes economic sense to implement and maintain.

Why It’s Needed

An AI system may be innovative and exciting, but if it does not offer a good Return on Investment (ROI) or if the market conditions are not favorable, it can quickly become a financial burden.

Components

• ROI (Return on Investment): Weighs economic benefits against costs.

• Market Risk: Evaluates the market factors that could affect the system's success.

Social and Ethical Concerns

The goal is to evaluate the broader societal and ethical implications of the AI system.

Why It’s Needed

AI technologies have the potential to impact employment, social interactions, and even ethical norms. These impacts need to be studied and mitigated to ensure responsible development.

Components

• Social Impact: Assesses the system's impact on employment, social interactions, etc.

• Ethical Alignment: Ensures alignment with societal values and ethical norms.

Legal Risks

Intent

The intent is to identify and assess the legal risks associated with the AI system, such as Intellectual Property infringement or legal liabilities.

Why It’s Needed

Legal consequences can be severe and costly. Being proactive in identifying and mitigating legal risks is crucial for the long-term viability of the system.

Components

• Intellectual Property: Assesses the risk of IP infringement.

• Liability: Considers legal responsibilities if the system fails or causes harm.

Conclusion

By integrating these additional dimensions into the original KAIRI framework, we offer a more comprehensive and holistic risk assessment methodology for AI applications.

This extended framework not only addresses technical and regulatory aspects but also incorporates crucial human, social, ethical, and legal elements necessary for the responsible development and deployment of AI systems.

References

[1] Artificial Intelligence risk measurement