AI Governance for “Trust from Society” in AI

Artificial intelligence (AI) is becoming increasingly important as a source of innovation and it was against this backdrop that the Ministry of Economy, Trade and Industry (METI) published its AI Governance Guidelines^* on July 9, 2021. I would like to start by talking about what led up to the publication of these guidelines.

AI offers considerable benefits for society. Image recognition AI, for example, is used in the development of autonomous driving and in the interpretation of X-rays. On the other hand, if discrimination or other forms of bias are inherent in the training data used to build AI systems, those biases may be exposed by the AI system. The history of work on AI governance is one of seeking to enjoy the benefits that AI brings while also dealing with its negative consequences.

To advance this work, the Japanese Cabinet Office published a set of high-level guidelines entitled “Principles of Human-centric AI Society” in March 2019. This document defined seven such principles, including privacy, security, fairness, accountability, and transparency. Debate within the government on AI governance for facilitating work that follows these high-level guidelines began in earnest in June 2020. METI released a document entitled “AI Governance in Japan” in January 2021 and called for public comment. The METI considered the feedback received and published AI Governance Guidelines in July.

In this work, METI defines “AI governance” as “design and operation of technological, organizational, and social systems by stakeholders for the purpose of managing risks posed by the use of AI at levels acceptable to stakeholders and maximizing their positive impact.” I would like to explain this further in the context of social systems.

Looking at Figure 1, the values sought after by society and embodied in the AI principles are listed under the “What” heading. The current challenges are shown under “How.” In practice, there are a variety of ways in which we could go about fulfilling the social principles of AI. One possibility would be to introduce industrywide (horizontal) rules that focus on the distinctive functions of AI. We could also establish technical standards. Yet another option is to introduce rules specific to particular sectors such as automobiles or healthcare. The challenge for AI governance is how to combine these in a way that maximizes the positive impacts of AI while minimizing the negative ones. Having considered industry-wide rules, METI has concluded that it is better to issue guidelines that are not legally binding. This led to the formulation of our AI Governance Guidelines.

The METI AI Governance Guidelines provide guidance on how to implement AI governance at the corporate level. The guidelines also draw on METI’s agile governance framework. This framework applies to the establishment of governance practices in fields such as AI where the technology is still evolving and the nature of public acceptance is still changing.

Figure 2 shows AI governance as a nested loop. It begins with a condition and risk analysis. In the case of AI, what matters is to understand the technology, public acceptance, and one’s own circumstances. This analysis is then used as a basis for setting goals. This involves determining which values are most important to the companies concerned. After this, the next step is to establish AI management systems for achieving the goals. To achieve these goals in the case of AI, it is important to conduct an as-is/to-be analysis to determine how the current situation differs from the goals. Also important are to improve literacy and to encourage businesses to work together. The implementation shown at the bottom of the loop relates to monitoring of the AI management system and individual AI systems. Appropriate external disclosure is also important. The evaluation in the inner loop looks at whether the AI management system is functioning as it should, and, if not, how it should be improved. Furthermore, if the business environment changes significantly, it will become necessary to perform a new condition and risk analysis and to repeat the process from the goal-setting step. The AI Governance Guidelines are designed to allow companies to engage in continuous improvement through dialogue with stakeholders. As one of the people involved in their formulation, I look forward to seeing these guidelines put into practice, even if only gradually.

Hitachi has also started to take action on AI governance. Recognizing that AI is such a rapidly moving field, the question of how to establish governance practices that are both flexible and fast is of particular importance. Today I would like to talk about what Hitachi is doing in this area in the context of METI’s agile governance framework.

First of all, in terms of the goals in Figure 2’s nested loop, Hitachi has established its own principles for the ethical use of AI that draw on what we learned from working through the analysis of AI situation and risk. One feature of these principles is that the setting of goals should be informed by an understanding of the AI lifecycle. Drawing on our considerable experience of working in areas such as social infrastructure that require long-term operation, Hitachi has set criteria for action that take account not only of those items to be addressed that apply generally, but also to the separate planning, societal implementation, and maintenance and management phases.

The next step after this is to establish the organization, people, and practices needed to operate the system design that in turn provides the governance to achieve the goals that have been set (see Figure 3).

The Lumada Data Science Laboratory (LDSL) is an organization within Hitachi that employs data scientists. In terms of specific measures, Hitachi has brought together a range of experts to form a specialist AI ethics team within the LDSL and has es-tablished mechanisms for coordinating with other groups across the company, including our privacy protection advisory committee. An AI Ethics Advisory Board made up of outside experts has been established to accumulate knowledge as well as to ensure objectivity. In addition to working in partnership with this advisory board, the AI ethics team functions as a center of excellence (CoE) with a role that encompasses information sharing, education, and training.

Operationally, the work involves the assessment of AI ethical risks, the conducting of evaluations based on checklists and by the AI ethics team and advisory board, and the implementation of actions and other measures identified on this basis. Based on the concept of “AI ethics by design,” Hitachi is seeking, wherever possible, to achieve the goal of reviewing and acting on its principles guiding the ethical use of AI from the planning and design stage or from the proof-of-concept (PoC) stage.

Finally, the evaluation in the inner loop involves periodically taking time out to review the system (at intervals of about six months) and to make improvements besides day-to-day operations. To achieve this, the AI ethics team as a CoE engages in continuous learning on topics such as policies and technologies that relate to AI and ethics.

Hitachi’s Social Innovation Business deals with social infrastructure and services with a strong public interest component, extending from health, medicine, and care to transportation and finance. Accordingly, being aware of just how much its activities can impact society, Hitachi has an obligation to take ethical factors into account in the development and utilization of AI in its businesses. Today I would like to tell you a little more about the practical steps Hitachi has been taking in this area.

Figure 4 shows an example of how Hitachi goes about technology governance. Further details can be found in the article “Principles and Practice for the Ethics Use of AI in Hitachi’s Social Innovation Business.” Among the key factors emphasized at Hitachi are: (1) Establishment of in-house governance practices, (2) Training of staff to develop and use AI correctly, (3) Dialogue and consensus-building with stakeholders, and (4) Research and development of technologies that underpin AI ethics. To provide a common set of guidelines and terminology that can be used in these activities, Hitachi has also formulated its own principles guiding the ethical use of AI in its Social Innovation Business.

Without a deep understanding of their technical characteristics and the changes they impose on society, new technologies such as machine learning and other forms of AI are at risk of being misused. Accordingly, a deep understanding of AI itself is a prerequisite for putting AI ethics into practice and this is something that needs to be addressed systematically and over the long term.

Hitachi has experience engaging in dialogue and consensus-building with stakeholders. This has included the establishment of the Hitachi-UTokyo Laboratory at the University of Tokyo and our work in partnership with Chulalongkorn University in Thailand. Recognizing the importance of talking to and coordinating with multiple stakeholders in the research and practical deployment of AI, Hitachi is stepping up its efforts in this area.

Hitachi has also established an AI Global CoE for the research and development of AI at facilities located across Japan, Asia, North America, and Europe. This involves going about research and development in an inclusive manner that takes diverse viewpoints into account. AI ethics covers a range of aspects, including safety, fairness, transparency, security, and privacy, and Hitachi’s research work is likewise broad in scope. In one example, Hitachi is working in partnership with Stanford University on world-leading research into techniques for enhancing fairness using explainable AI (XAI). Along with research into AI itself, other work is looking at data governance and at trust governance. One such example involves the development of a system for optimizing the deployment of emergency services that use XAI. Services with a strong public interest component require high levels of fairness and explainability. Accordingly, it is by utilizing techniques designed to meet this need that successful practical implementation has been achieved.