Work

Operationalizing Transparency and Explainability in Artificial Intelligence through Standardization

Year: 2022
Type: dissertation
Abstract: As artificial intelligence (AI) has developed, it has spread to almost every aspect of our society, from electric toothbrushes and telephone applications to automated transportation and military use. As AI becomes more ubiquitous, its importance and impact on our society grow continuously. With the pursuit and development of more efficient and accurate artificial intelligence applications, AI systems have evolved into so-called “black box” models, where the operation and decision-making have become immensely complex and difficult to understand, even for experts. As AI is increasingly applied in more critical and sensitive areas, such as healthcare, for instance in support of diagnoses, the lack of transparency and explainability of these complex models and their decision-making has become a problem. If there is no understandable argumentation backing up the results produced by the system, its use is questionable or even ethically impossible in such areas. Furthermore, these AI systems may be misused or behave in very unexpected and potentially harmful ways. Issues related to the governance of AI systems are thus more important than ever before. Standards provide one way to implement AI governance and promote the transparency and explainability of AI systems. This study sets out to examine how the role of standardization in promoting AI transparency and explainability is perceived from an organizational perspective and what kind of AI transparency and explainability needs are identified among different organizational actors. In addition, efforts will be made to identify possible drivers and barriers to the adoption of AI transparency and explainability standards. The research has been carried out by interviewing representatives from a total of 11 different Finnish organizations working in the field of AI. The data gathered from the interviews has been analyzed using the Gioia method. Based on this analysis, five different roles for standards were identified regarding the promotion of explainability and transparency in AI: 1. Facilitator, 2. Validator, 3. Supporter, 4. Business enhancer, and 5. Necessary evil. Furthermore, the identified AI transparency and explainability needs are composed of the needs for ensuring general acceptability of AI and risk management needs. Finally, the identified drivers for adopting AI transparency and explainability standards comprise the requirements of the operating environment, business facilitating drivers, and business improvement drivers, whereas the barriers consist of the lack of resources, lack of knowledge and know-how, downsides of standardization, and incompatibility of standardization and AI. In addition, the results showed that the implementation of possible standards for AI transparency and explainability is largely driven by binding legislation and financial incentives rather than ethical drivers. Furthermore, building trust in AI is seen as the ultimate purpose of transparency and explainability and its standardization. This dissertation provides an empirical basis for future research regarding the need for AI standardization, standards adoption, and AI transparency and explainability from an organizational perspective.
Author: Tamminen, Panu
Language: English
Open Access status: green
OpenAlex ID: W7023314487
Cites: 75
Cited by: 4
Related to: 10