How to build AI that society wants and needs

While AI holds great promise for society, the speed of its advancement has far outpaced the ability of businesses and governments to monitor and assess the outcomes properly. The relative ambiguity of regulatory oversight throughout the world prevents AI from directly reflecting society’s needs. It is important that organizations take steps to enable and showcase trustworthiness to all stakeholders and build the reputation of the organization’s AI.

Trust in AI starts with stakeholders understanding that a particular organization uses AI responsibly. It is unlikely that external stakeholders will identify individual AI systems as “trustworthy” or “untrustworthy”; rather, an organization is considered trustworthy or not and AI systems inherit the organization’s reputation. In the same way that an organization’s human staff showcases the organization’s values, the behaviours of the AI system are both a manifestation of and an influence on the organization’s reputation.

Training staff is a familiar challenge to most organizations, but the challenges of implementing ethical and trustworthy AI are new and different. They are, however, well documented with more than 90% of surveyed companies reporting some ethical issues. How can an organization do better?

In the current ambiguous regulatory environment, the complexity of AI systems drives organizations to seek new means to support their AI development. Most sophisticated technology companies, at a bare minimum, indicate they have the fairness, ethics, accountability and transparency (FEAT) principles at the centre of AI development and more than 170 organizations have published AI and data principles. However, many organizations tend to “build until something breaks” without consideration for who is affected by the break.

We believe that tangible progress toward the responsible use of technology can be made by taking advantage of people, process and technology to bridge these gaps. Here’s how:

Organization leaders – from managers to the C-suite and board of directors – often have little understanding of the assumptions and decisions made throughout the development and implementation of AI. Regardless of leaders’ understanding of the AI, they own the reputational and financial outcomes (both positive and negative). Data scientists, on the other hand, can find it challenging to take all the guidelines, regulations and organizational principles into account during the development process.

In both cases, the challenge is not generally a lack of understanding of what it means to be responsible; it is a lack of insight into what factors are important at different levels of the organization and how they affect outcomes. Finding ways (processes and tools) to bring disparate groups together can be highly effective.

Stakeholders expect AI to be overtly fair, aspiringly ethical, accountable and as transparent and open as possible. To maximize their external stakeholders’ trust, a systematic and visible process must be used to develop all responsible technology.

Guidelines for the development of responsible technology are plentiful and varied. It can be a challenge to know which to adopt. An organization must pick one and use it. It is better to be systematic and have a framework that is 80% aligned with your needs than no framework at all.

When using a framework for the development of responsible technology, there must be some form of artifact (documentation or report) that can be shared with stakeholders. The process is then systematic and visible. These artifacts serve both as proof that a process was implemented and followed, and as a mechanism to guide iteration and feedback.

Tools must be developed and adopted that can efficiently and effectively integrate these processes into AI development and deployment. Such tools must facilitate the employment of guidance frameworks, expedite analyses that assess potential problems like bias and generate useful artifacts for internal and external stakeholders.

AI developers need objective tools to help assess their work and allow organizational stakeholders to provide feedback. Organizational leaders need standardized assessments of the AI’s overall risk and financial metrics like return on investment (ROI). Additionally, tools that explain how AI systems make decisions can enable executives and boards of directors to have better oversight and understand risks and any financial impacts. These tools can provide a standardized structure and artifacts that benefit both data scientists and organizational leaders and help overcome the disconnect between the requirements to create and the requirements for ethical outcomes.

To enable trust in an organization’s AI, people, processes and tools must allow internal stakeholders to build responsible technology. Without these concrete steps, the “build until it breaks” mentality will likely cause a breach of trust that could derail the organization. Empowering an ethical and transparent process, where colleagues and managers hold each other accountable and mutually enable desirable behaviour is how ethical people foster an ethical culture for responsible AI development.