What steps to take if you want to regard fairness in your AI system development
Introduction
Why fair AI systems matter for your organisation
Artificial Intelligence systems are being integrated into more and more applications in different use contexts. At the same time, societal awareness of the potential harms that such systems might cause has also risen. Heightened societal expectations and backlashes against organisations that develop or deploy AI systems that are perceived to be unfair, as well as new regulations such as the AI Act, can be expected in the near future. Building fair AI systems will therefore become crucial to ensure the trust of users and the success of the business.
The processes for developing and maintaining fair AI systems will also make them higher quality products: Fairness requirements often include user-centricity, and will therefore target a larger and more diverse group of potential users. Additionally, system robustness and security will benefit from the increased testing and monitoring; while the documentation and transparency requirements should also render the systems easier to manage and maintain, as well as for end-users to use and understand. Finally, it is essential to note that AI systems often discriminate because they underperform for certain groups of people. This is a product quality issue that can cause loss for the company, the removal or mitigation of such defects can open up new markets and create new customers for the business.
What is fairness in the context of AI systems?
For whom and what is this process model?
Basic structure: Technical Management Processes vs Technical Processes
Fairness in the context of AI systems can have several meanings and needs to be defined in the use context of the AI systems. Other aspects of trustworthiness, such as human agency and oversight, technical robustness and safety, privacy and data governance, transparency, diversity, societal and environmental well-being and accountability may be mentioned when relevant for fairness but are not the focus of the fAIr by design process model. (see also: ETHICS GUIDELINES FOR TRUSTWORTHY AI High-Level Expert Group on Artificial Intelligence, https://ec.europa.eu/newsroom/dae/document.cfm?doc_id=60419)
Whereas discrimination is precisely defined in jurisdiction, fairness is very subjective and differs highly between different people, cultures, backgrounds, regions, context of use and organisations.
The current process model was developed by the research consortium fAIr by design. It is intended for anybody who is interested in improving fairness of AI systems. It offers support in developing, deploying and maintaining fair AI systems and can be used in different contexts and for different technologies and uses. Adhering to the process model should offer help in being prepared for future standards and regulations, as well as facilitating eventual third-party audits; but it does not give any guarantee of compliance.
The process model includes steps for all stages of AI system development and deployment, for an interdisciplinary team.
In the process model tasks and project activities will be presented in two different process categories: the technical management processes and the technical processes, consistent with ISO 15288 and ISO 5338 terminology. Both process categories are highly interwoven and influence each other, but from the system life cycle perspective, they address different aspects: the technical management processes can be thought of as “horizontal”, spanning several life cycle stages; whereas the technical processes are more “vertical”, and consist of activities that take place within one life cycle stage. Thus, for better understanding and easier task allocation, the two process categories are presented separately.
Technical management processes
The Technical Management Processes are used to manage and control the system development and the allocated resources. The technical management processes can be thought of as “horizontal”, spanning the entire system development life cycle, and ensuring the system satisfies objectives at all stages of its life cycle. They form the basis for the technical processes and enable an efficient workflow and beneficial framework conditions (such as terms of costs, planning of the project and the overall management of the project and team). The Technical Management Processes are identical to the technical management processes identified in ISO 15288 and ISO 5338
Technical processes
The Technical Processes relate to technical actions that need to be taken at particular stages of the AI system life cycle. They provide specific outputs that are used to build, evaluate, or maintain the AI system, and consist of concrete and practical steps to be taken in order to develop and use the AI system as planned. The technical processes are identical to the processes defined in ISO 5338.