Page 19 - U4SSC Guiding principles for artificial intelligence in cities

• Identifying mechanisms to audit AI technology deployments and continuously monitor their
deployment;
• Incorporation of unbiased diverse datasets;

• Ensuring equal and broad representation of the target population (or potential group) for which
AI systems are designed;
• Implementation of accessible and user-friendly AI systems to realize the human rights of people
with different abilities (mitigating barriers to access and usability where needed);

• Proactive assessment of existing non-AI biases, inequalities, and discriminatory practices to
avoid their ratification into AI systems;
• Sharing the benefits of AI systems and empowering people equally;

• Ensuring diverse participation and engagement during AI systems design; and

• Designing and deploying AI systems by staff who are responsible and equipped with adequate
skills in order to not introduce further biases.

3.2.4 Explainable and transparent

AI systems use sophisticated technologies and complex mathematical algorithms. The rationale
behind the actual decisions and the results of AI systems are not easy to discern since they tend
to be opaque in most cases due to their inherent design.

On the other hand, AI systems today make a large spectrum of decisions ranging from relatively less
impactful ones to even life and death situations. Consequently, it is important to understand how
AI systems reach their decisions and results in order to make sense of them, and more importantly,
to enhance them.

Hence, this principle allows cities to procure, develop, deploy, and use explainable and transparent
AI systems.

Implementation Considerations: Cities can adopt various mechanisms to help enhance the
explainability and the transparency of their AI systems. These mechanisms include:

• Identifying appropriate use cases for using open-source algorithms (noting the potential risks,
such as cybersecurity risks associated with open-source algorithms.);
• Conducting algorithm audits and AI systems’ independent verification;

• Implement extensive testing frameworks whereby different data set modalities are used to
check for potential bias. Creating public AI or algorithm registers (e.g., Amsterdam algorithm
register, Helsinki AI register);
• Informing users when and how AI makes decisions for individuals or for important city matters
to enhance transparency;

9