Artificial Intelligence Ethics Principles

We’ll develop and use AI in a manner that minimizes the creation or reinforcement of unfair, biased and discriminatory impacts

Our transparent AI development processes will prioritize fairness. We’ll endeavor to use training datasets that reflect the diversity of society to eliminate algorithmic bias that places unprivileged and underrepresented groups at systematic disadvantage. We’ll judge fairness based on the same standards as decision-making processes by humans and work to detect and eliminate fairness-related harm within our AI.

We’ll apply strong testing, safety and security practices to minimize unintended results or outputs from our AI systems and to include appropriate human oversight capabilities

We’ll develop AI using our secure product development lifecycle process. We’ll regularly test and maintain our AI systems with the goal of ensuring that they operate reliably, safely and consistently under normal circumstances and in unexpected conditions. We’ll share relevant information about inherent limitations and error rates of the system with users of our AI systems for mitigating actions when appropriate.

We’ll develop our AI so that systems can make decisions autonomously only if they operate within the basis of clearly defined principles. This will limit their decision-making powers and help ensure the safety of our users, employees, partners and other stakeholders.

We’ll be clear about when, why and how we deploy our AI systems and will help our customers and partners do the same

We believe that transparency increases trust and confidence in our AI systems. When appropriate, we’ll provide users with information and training related to the enterprise use of an AI system, the intended purpose of such AI system, and the types of data used by the AI system.

For AI models whose outcomes may materially affect the rights or safety of persons, we’ll generate AI model cards that provide developers, users and other stakeholders of our AI systems information about the development and deployment of various AI models.

The model cards will give a useful, non-technical explanation of the behavior of our AI systems and their components. Stakeholders can then comprehend how and why they function and can identify and mitigate potential safety, privacy, bias or other issues.