The field of artificial intelligence (AI) has witnessed significant advancements in recent years. At least since the release of the chatbot ChatGPT, AI has entered the public consciousness. From everyday products of tech companies, such as Google’s AI-enhanced search engine, the language translation service DeepL, and Meta’s AI assistants for social media to domain-specific applications including disease diagnosis, fraud detection, industrial quality control, precision farming, autonomous vehicles, and personalized product recommendations, the impact of AI on our lives and work is profound. However, as the use of AI becomes more prevalent, the number of reports on failures and misuse of this technology is increasing, and discussions on the potential risks and ethical implications are becoming more frequent. A number of incidents have been documented in which AI systems have caused harm. For example, AI-based chatbots and deep fakes have contributed to the spread of misinformation, even having the potential to influence electoral processes. AI-based job application systems and facial recognition systems have been shown to discriminate against specific population groups. Autonomous vehicles that were unable to demonstrate an appropriate reaction to unusual situations have been implicated in severe accidents. The illustrative examples demonstrate that the deployment of AI entails a number of risks. As an answer, several countries have already initiated the process of formulating guidelines and regulations with the objective of ensuring the socially beneficial and responsible use of AI. For instance, in 2021, the European Commission proposed the AI Act, which establishes regulations for AI systems based on their risk level. Likewise, in the research community, several fields have emerged that address specific aspects of AI risks. For example, explainable AI (XAI) aims to make AI systems more transparent and interpretable, while the goal of fair machine learning (Fair ML) is to create AI systems that are free from bias and discrimination. In recent years, interdisciplinary fields have emerged with the aim of integrating these disparate aspects. In this context, trustworthy AI (TAI) has attracted a growing level of attention. TAI investigates approaches to create trust in the development, deployment, and use of AI systems. The field delineates requirements, predominantly from a technical perspective, that are recommended to be considered at all stages of the AI life cycle to guarantee that AI systems are dependable, transparent, and equitable. The definitions of trustworthiness and its requirements are varied, whereas a particularly noteworthy set are the seven requirements defined by the European Commission’s High-Level Expert Group on AI (HLEG) in 2019: (1) human agency and oversight, (2) technical robustness and safety, (3) privacy and data governance, (4) transparency, (5) diversity, non-discrimination and fairness, (6) societal and environmental well-being, and (7) accountability. In the field of TAI research, a number of activities have been undertaken with the objective of defining principles, dimensions, or requirements for trustworthiness. Another research stream has commenced to investigate the interdependencies and conflicts between different requirements, that must be considered when aiming to enhance the trustworthiness of AI systems. Furthermore, efforts have been made to create frameworks and tools to support the development of TAI systems. However, the understanding of trustworthiness and its requirements is still evolving, and the existing practical support is often abstract without concrete implementation guidelines. This dissertation aims to build upon existing research on TAI and to contribute to the ongoing discourse surrounding trustworthiness in AI. A main focus lies on the practical implementation of trustworthiness throughout the entire AI lifecycle. The dissertation comprises four largely independent studies that address trustworthiness at different stages of the lifecycle, using examples from two distinct application domains. The following sections provide a brief overview of the studies. Study 1 builds upon existing research in the field of trustworthiness in AI by conducting a systematic literature review on recent literature on TAI between 2018 and 2023. Given the considerable number of reviews and surveys in this field, which can be attributed to the multi-layered nature of the topic, a form of meta-review approach combining quantitative and qualitative analysis is deemed appropriate. The meta-review offers an overview of the definitions of trustworthiness and the requirements for TAI, as well as methods for evaluating and enhancing trustworthiness. The findings suggest that there is a multitude of interpretations of trustworthiness and its associated requirements. However, fairness, robustness, privacy, and explainability emerge as the most frequently referenced aspects. The current approaches for evaluating and improving trustworthiness tend to focus on individual requirements, with less attention paid to their interdependencies. In summary, the results underscore the intricate and relatively immature state of the field. Study 2 employs the use of facial analysis as a case example to experimentally investigate the viability of developing an evaluation framework for assessing the trustworthiness of machine learning (ML) models. The framework consolidates the four principal trustworthiness requirements - fairness, robustness, privacy, and explainability - into quantifiable scores. The results of applying the framework on four different datasets demonstrate the feasibility of evaluating the trustworthiness of ML models and the high dependability of the results on the data. They further show that a single score for trustworthiness is not realizable. However, the metric scores and aggregated dimension scores for each requirement provide informative insights into the trustworthiness of the models, while offering a less complex representation. Study 3 employs an experimental methodology to examine the influence of selected training adaptations meant to enhance either fairness, robustness, privacy, or explainability on the respective other dimensions and model accuracy. To this end, the study employs the context of facial analysis and the evaluation framework developed in Study 2. The objective is to evaluate the interdependencies between the requirements and to quantify the extent of potential conflicts between them. The findings confirm that enhancements to the overall trustworthiness are difficult to achieve, as improvements in one dimension can have diverse effects on the other dimensions. In our use case, the improvement of robustness through data augmentation is the most synergistic approach, while fairness, privacy, and explainability enhancements can have pronounced negative effects on model accuracy and the other dimensions. Study 4 sheds light on the trustworthy operation of an ML system in the context of industrial manufacturing. Guided by the design science research (DSR) methodology, the study designs and implements a trustworthiness testing and monitoring solution for an ML-based optical inspection system used during a laser welding process in electric drives production. The testing and monitoring solution is designed to be integrated into the existing MLOps pipeline to enable a continuous training and deployment of ML models. The testing and monitoring results are presented in dashboards that offer insights into the ML models’ trustworthiness. The evaluation of the solution demonstrates its feasibility and potential to detect performance issues in unconditional situations, as well as data drift and anomalies in production. In summary, the dissertation contributes to the field of TAI by providing insights into the practical implementation and quantification of trustworthiness in AI systems along different stages of the AI lifecycle, from development to operation. The performed studies underline the substantial value of experimental research in the area of TAI, discuss challenges of developing and operating TAI systems, as well as reflect on the transferability of results within the field of TAI.
A. F. Schreiner (Fri,) studied this question.