How do you know if an AI system can be trusted?

1 What is AI trustworthiness?

AI trustworthiness is the degree to which an AI system can be relied on to perform its intended functions, respect human values and rights, and contribute to social good. AI trustworthiness is not a fixed or static property, but rather a dynamic and context-dependent one. It depends on the design, development, deployment, and evaluation of the AI system, as well as the expectations, perceptions, and interactions of the users and stakeholders.

2 Why is AI trustworthiness important?

AI trustworthiness is important for several reasons. First, it affects the acceptance and adoption of AI systems by users and stakeholders. If an AI system is perceived as trustworthy, it can enhance user satisfaction, loyalty, and engagement. If not, it can lead to user frustration, distrust, and rejection. Second, it impacts the outcomes and impacts of AI systems on individuals, groups, and society. If an AI system is trustworthy, it can support decision making, problem solving, and innovation. If not, it can cause errors, harms, and risks.

3 How to measure AI trustworthiness?

When assessing AI trustworthiness, there is no single, universal metric or method. However, there are some general principles and criteria that can be used as a guide. These include accuracy, transparency, explainability, accountability, fairness, privacy, and security. Accuracy refers to the extent to which an AI system produces correct and consistent results. Transparency looks at the extent to which an AI system reveals its inputs, outputs, processes, and logic. Explainability examines the extent to which an AI system provides understandable and meaningful reasons for its actions and decisions. Accountability considers the extent to which an AI system and its developers, operators, and users are responsible and liable for its performance and consequences. Fairness evaluates the extent to which an AI system treats all users and stakeholders equally and impartially. Privacy looks at the extent to which an AI system protects the personal data and identity of users and stakeholders. Finally, security assesses the extent to which an AI system prevents unauthorized access, manipulation, and attacks.

4 How to improve AI trustworthiness?

Improving AI trustworthiness is not a one-time or one-sided task, but rather a continuous and collaborative process that involves multiple steps and actors. Designers must ensure the AI system is aligned with user and stakeholder needs, values, and goals, as well as ethical and legal standards. Developers must guarantee the AI system is built with high-quality data, algorithms, and tools, then tested for accuracy, robustness, and reliability. When deploying, it's important to integrate the AI system with the user and stakeholder environment, context, and expectations, while also monitoring for performance, feedback, and improvement. It's also necessary to evaluate the AI system for its outcomes, impacts, and risks; auditing for its transparency, explainability, accountability, fairness, privacy, and security.

5 How to trust AI trustworthiness?

Trusting AI trustworthiness is not an easy decision, but rather a complex and subjective judgment that depends on the characteristics and preferences of users and stakeholders, as well as the purpose of the AI system. Evidence, experience, reputation, emotion, and context can all influence trust in AI trustworthiness. Evidence refers to the availability and quality of data that supports the system's trustworthiness. Experience is based on the frequency and intensity of interactions with the system. Reputation is based on opinions and ratings from peers, experts, or media. Emotion is based on feelings and attitudes towards the AI system. And context refers to the situation and environment in which it is used.

6 Here’s what else to consider

This is a space to share examples, stories, or insights that don’t fit into any of the previous sections. What else would you like to add?