Foundations of trustworthy AI: Guiding the workforce
As an organization expands its use of AI, its teams should build individual AI solutions that are trustworthy from the start. This requires guiding the workforce so that everyone involved in creating the solutions understands the tenets of trustworthy AI, knows how to apply them, and makes them an integrated part of their work as they plan, create, operate, and revise each solution. This process has two main elements: education, so that each person knows what it takes to make AI trustworthy from the perspective of their own role; and specialized, human-centered Enterprise Design Thinking, to ensure that the design and planning of the solution incorporates what is needed for trustworthiness from the beginning.
The full workforce involved in an AI solution needs to have current, specific knowledge of how issues of trustworthiness — explainability, fairness, robustness, transparency, and privacy — affect their role and how to implement a trustworthy solution. Designers need to understand how to plan for trust concerns and to define use cases and processes that incorporate those concerns. Data scientists need to have a clear idea of how trust requirements influence their choices of data and methods. (For example, a protected attribute might be inappropriate to use as a feature that affects a model outcome, but its availability might be critical for measuring fairness.) Machine learning engineers need to understand the trustworthy requirements for Q/A, deployment, and operational monitoring. Development leads need to incorporate governance and transparency into the full projects. And decision makers in all areas need an overall understanding of what it takes to trust AI and how that affects their projects and practices.
At IBM, we have encapsulated our deep AI expertise into courses and learning paths, with a strong focus on the tenets of trustworthy AI. We have education around AI bias detection and mitigation, AI transparency and AI governance, leveraging open-source libraries for Trusted AI we have made available to the world, as well as our Watson technologies for managing and monitoring AI in production.
In addition to education, workforce guidance means that AI projects must explicitly incorporate planning around trust into their designs and plans from the start. Enterprise Design Thinking provides an ideal way to meet this goal when it is specialized to focus on AI and its trustworthiness. Enterprise Design Thinking hinges upon structured brainstorming and group planning exercises to plan project scope and has earned its popularity in software development and technology in general by encouraging creative thinking while grounding plans in specific business needs and supporting the continuous refinement used in agile development.
With a focus on enterprise-level AI goals, this same approach can drive key insights for AI strategy and practices across an organization. Likewise, with a focus on the key elements of performing AI and ensuring its trustworthiness, it provides the basis for planning AI lifecycle activities, from building the model to validating it and monitoring it in operation. These specialized exercises cover:
- Core AI content, such as characteristics of data sets, specific casting of the AI problem, and the way the AI task fits into the overall business workflow. These core elements expand to include privacy requirements around the data sets and use of outcomes in the workflow.
- Accuracy validation requirements, including robustness: resistance to drift (changes in the data or its implications over time) and adversarial input
- Fairness requirements, such as treating all groups equitably, how to identify those groups, and how this requirement interacts with privacy
- Explainability requirements, including who will be using the explanations and what constitutes an adequate explanation for the purpose
- Governance and transparency requirements, including processes that require approvals and information that must be tracked as a solution is compiled, such as characteristics of data sets and choices of algorithms
At IBM, we have defined Enterprise Design Thinking to address the range of needs as organizations build out AI strategies and operationalize and broaden their use of AI, in individual projects or across the enterprise. The expanded use of Enterprise Design Thinking has brought us to develop a new methodology named Enterprise Design Thinking for Data and AI that solves for data and AI problems and is an extension of the traditional Enterprise Design Thinking. We employ these as needed for each individual situation, to meet you where you are in your journey to transformation with trustworthy AI.