Towards Trusted AI: A Blueprint for Ethics Assessment in Practice

As evidenced by the information presented above, all countries except for the EU AI Act have voluntary regulations or soft-laws when it comes to AI systems. The guidelines often focus on the same principles with security being at the forefront. Privacy and protection are always among the principles, but their understanding differs between countries. As an effect, different court outcomes might appear. In the Indian case mentioned above the court decision was favoring the actor, but in a similar incident in the U.S., when Scarlett Johannson wrote to OpenAI about illegally using her voice, the company stopped the use of her persona, but on a legal level no measures have been taken [26]. This shows that AI governance and ethical frameworks vary across the globe in regard to regional, legal and cultural values, and even more when it comes to strategic interests in shaping digital power.

There are three competing regulatory models, each reflecting a different approach for the digital economy. The United States adopts a market-driven model, focusing on flexible frameworks, China follows a state-driven approach, emphasizing control, security, and social stability in AI development, and the European Union takes a rights-driven stance, prioritizing ethical standards [4]. These three distinct models – market, state, and rights-driven – illustrate that the global landscape of AI ethics is not only a mere reaction of technological advancements but also a manifestation of the underlying political, economic, and cultural dynamics that concretize each region’s approach to AI governance.

In summary, a global ethical framework, with the objective to guide the deployment of trusted AI and to promote the responsible use of AI, implies the need of a process blueprint. The blueprint for an AI ethics assessment must fulfill two acceptance criteria. The first criterion refers to its high level of independence, which implies it is agnostic to the underlying regulatory model, to the deployed AI algorithm, to the technology in which the AI model is embedded in, and it is agnostic to the needs of the industry sector or to the business model or scale of business. The second criterion of the blueprint allows for adaptivity to varying comprehension of ethical principles and values. As already been pointed out, the interpretation or choice of ethical principles depends not only on the cultural perspective, but it is also tailored to specific industry needs and it also aims to maximize the space for AI innovation for which most national AI strategies of countries define a leading position. Lastly, the AI ethics process blueprint that fosters a trusted AI ecosystem cannot be static. The blueprint itself requires a review and update process that adapts to advancements in AI.

While most AI assessment solutions comprise high-level ethics principles and evaluation tools [7], [20], they miss the practical aspects needed for operationalization in the cycle from idea-to-AIOps deployment. Therefore, our aim is to build a generic AI Ethics Assessment Blueprint for the evaluation of the entire lifecycle of an AI system, from design and development to deployment.

The Blueprint’s adaptable framework integrates ethical principles and their associated assessment tools as inputs, leading to a materiality analysis of the AI system. To achieve our goal, we utilized the UNESCO Ethics Principles [32] and the UNESCO Ethical Impact Assessment Tool [33]. We chose the UNESCO ethics framework for two reasons, first, it is congruent with the EU definition of trustworthy AI and, second, it is a global reference standard, adopted by all 193 UNESCO member states in November 2021. An overview of the UNESCO Ethics Principles is provided in appendix A.

The Blueprint for AI Ethics Assessment serves as a facilitator, ensuring that the development process and lifecycle of an AI system are supported rather than constrained. It is designed to enhance and ease the ethical evaluation process, but also to support the ethical and responsible design, development, and deployment of AI systems, providing a structured approach that does not hinder the AI system different lifecycle phases.

In accomplishing operationalization, an AI ethics assessment framework must contain at least the following three components: (i) high-level ethics principles, (ii) an ethics assessment tool corresponding to the ethics principles, and (iii) a set of evaluation measures relating to key performance indicators (KPIs).

Ethics metrics or their defined thresholds provide an important instrument in the decision-making process, for example in selecting mitigation strategies as part of the results of an ethics assessment. Without outcome-driven ethics metrics along the AI lifecycle pathway the operationalization of an ethics assessment framework remains a challenging milestone. To solve the challenge, we propose to develop a phased approach which is described in the next section.

The decision to implement a five-phase process in the Blueprint for AI Ethics Assessment is rooted in the need to establish a structured approach to addressing ethical challenges throughout the entire lifecycle of an AI system. This phased approach was chosen to ensure that ethics are not treated as an afterthought or a box-ticking exercise but are an integrated part of AI development and deployment. Because AI technologies present complex and multifaceted ethical dilemmas that require ongoing, context-sensitive assessment, a single-stage process would be insufficient to capture the nuanced and evolving nature of the ethics issues.

The five phases are based on two motivational drivers: first, to reflect the ethics principles and second, to incorporate the technicalities of the software-engineering needs. Furthermore, focusing on concrete phases such as system design, data management, model development, system monitoring, and human-centric evaluation, the Blueprint can ensure that ethical issues like fairness, transparency, and accountability are assessed in relation to the actual system development process.

Each of the five phases corresponds to a distinct aspect of the AI system’s development, from design to human-centricity evaluation, allowing for a step-by-step integration of ethics in an iterative manner, with each phase building upon the previous one. The rationale behind dividing the process into five phases is to break down the complexities of AI ethics into manageable components, each targeting specific risks and challenges that might emerge at different stages of the AI lifecycle. This phased approach enables continuous feedback loops, ensuring that ethical compliance is not static but evolves alongside the system itself, thus creating a more dynamic and responsive framework. A detailed description of each phase will be presented below.

Our framework addresses the three main first level stages of an AI system lifecycle - the system design stage, the development stage, and the deployment stage. On the second level, we define five phases where each phase entails two processes on the third level and one result as outcome. Every process includes methodologies and practices aimed at addressing the specific needs and challenges of its corresponding stage within the AI system lifecycle. A schematic of the Blueprint for AI Ethics Assessment in Practice is shown in figure 2.

Below, the detailing out the five phases of the Blueprint are presented:

The blueprint is designed to enable the operationalization of AI ethics assessment. The applicability of the blueprint is manifold. It addresses AI systems working alongside human subjects, for example in robotic-assistance or in AI-supported decision-making. It assesses impact along the AI supply chain where downstream AI-driven products or solutions are built around a (generative) AI model offered by an upstream provider. To allow for diverse applications to be assessable, we established a harmonized terminology by mapping assessment questions of a chosen ethics framework to our phases and processes of the AI ethics assessment blueprint. Our approach focused initially on measures of the UNESCO ethical impact assessment with about 160 questions or measures, but it can easily be extended to other frameworks, for example to the European Commission’s Assessment List for Trustworthy AI (ALTAI). The mapping of measures is visualized in the dendrogram in figure 3. Answers to the ethics framework assessment questions will then contribute to the outcomes of the blueprint in practice. An effective and timely assessment will require a screening and application or use case specific selection of framework questions. It is not required to answer all questions to summarize in a meaningful outcomes report. To guide the screening process of questions for relevant outcomes a definition of the outcomes measures of the blueprint is presented in figure 3.

Given the diversity of AI use cases the blueprint can address we cannot define application specific output measures and instead we propose an outcomes category for each of the 5 phases. For each outcomes category we provide a set of selectable outcome measures which we denote as AI Ethics key performance indicators (KPIs). The AI Ethics KPIs will then ensure the responsible design, development, and deployment of AI systems. The following outcomes catalogue is exemplary and has no intention of being complete.

• $\mathrm{■}$

  Define governance: Assemble an Ethics Board, with roles, responsibilities, system objectives, and accountability structures defined within the first few months of project initiation.

• $\mathrm{■}$

  Identify potential incidents: Analyze current literature to identify potential incidents, and related proposed mitigation measures for the specific use case.

• $\mathrm{■}$

  Define review process: Establish regular review mechanism for ethics clearance by process of multi-stakeholder collaboration including ethics advisors.

• $\mathrm{■}$

  Define scope: Select AI system features that must undergo ethical screening by the Ethics Board.

• $\mathrm{■}$

  Identify at-risk-groups: Identify at-risk groups which may be systematically disadvantaged by the AI system.

• $\mathrm{■}$

  Define fairness: Select fairness objectives and associated fairness metrics to measure consequences of biased or unfair data on model outputs with respect to harms and benefits which (at-risk) individuals may receive by use of the AI system.

• $\mathrm{■}$

  Implement bias detection: Implement bias detection processes throughout the AI lifecycle at defined bias checkpoints based on selected fairness metrics. Definition of processes for mitigation of bias present in data or outputs that impact fairness of the AI system.

• $\mathrm{■}$

  Implement fairness audit: Define regular screening and data audits to ensure compliance with fairness data guidelines and ethics principles.

• $\mathrm{■}$

  Assess transparency: Document performance and uncertainty of the AI model with respect to fairness objectives. Justify personal attributes in the data that are used for fairness assessment.

• $\mathrm{■}$

  Assess explainability: Conduct explainability assessment of the AI system to ensure that the system’s potential decision-influencing processes are clear and understandable to stakeholders of the system and to users and addresses of the system’s outcome. Explainability assessments are of paramount importance for decision-assist systems. Here the focus is on detecting decision boundaries and deriving concrete recommendations for actions in gray area situations or for high-stakes decisions.

• $\mathrm{■}$

  Assess safety and security: Conduct a safety and security evaluation of the AI system to ensure all identified risks to fairness and operational safety are documented and classified by materiality prior to deployment.

• $\mathrm{■}$

  Identify AI materiality: Document all risks associated with the AI model’s materiality and categorize all impacts with respect to severity and likelihood. Identify mitigation strategies for each risk.

• $\mathrm{■}$

  Update materiality classification: Define process for post-hoc assessment or audit of the AI model’s materiality classification. Flag any newly ob-served risks and resolve in continuous system improvement initiatives.

• $\mathrm{■}$

  Define system monitoring: Define system monitoring and review process to detect abnormal operation of the AI system. Address all identified ethical, operational, and security risks so that potential system impacts are aligned with fairness objectives.

• $\mathrm{■}$

  Measure incident metrics: Track incident response metrics so that monitoring enables fast time to detect (TTD) and fast time to resolve (TTR).

• $\mathrm{■}$

  Define mitigation plan: Define fallback or mitigation plan in case of trigger events from system monitoring or review.

• $\mathrm{■}$

  Update mitigation strategies: Evaluate effectiveness of mitigation plans by measuring incidents after a post-implementation phase. Align updates of mitigation measures with new risks or evolving model behaviors.

• $\mathrm{■}$

  Guide safe and responsible use: Develop operationalization guidelines for the AI system. Implement AI literacy and ethics awareness program to en-sure that all stakeholders understand how to use and interact with the AI system considering ethics, and system limitations. Ensure that users of a collaborative AI system understand the embodied ethics under normal operation and ethical boundaries. Assess regularly (by user feedback or questionnaires) stakeholders’ ability to exercise human oversight over the AI system. Train developers and system users in recognizing and mitigating ethical risks.

• $\mathrm{■}$

  Evaluate human centeredness: Analyze by regular post-deployment reviews that human-AI interaction and human oversight remain effective. These reviews will track how effectively the system supports AI-assisted decision-making.

• $\mathrm{■}$

  Establish AI training for professional development: Ensure regular updates of the AI literacy and sustainability training. Adjust training programs based on explaining observed versus expected outcomes, on system improvements, user and stakeholder feedback, and on advancement in state-of-art and energy-efficient technologies. Ensure that employees acquire sufficient knowledge in developing, improving, deploying or using the AI system throughout the entire life cycle.

• $\mathrm{■}$

  Measure impact on social goals: Identify Social Development Goals (SDGs) also known as Global Goals adopted by the United Nations [35], societal benefits/social goals, or sustainability goals where the AI system can create impact on. Ensure regular screening so that the system’s social impact aligns with ethical standards and long-term social benefits, and that the environmental issues are mitigated through sustainable practices during system operations.

• $\mathrm{■}$

  Measure energy consumption: Measure energy consumption and related costs during training and inference stages. Identify options to minimize the system’s carbon footprint, for example by choosing a smaller (foundational) AI model or by effective finetuning. Compare effects of model hosting on premise, on cloud and on edge (device).

We propose an AI system, generative AI for Arts Education with the acronym StableArtists. The technical realization of the system is based on a custom-trained text-to-image AI model that generates images based on a given prompt or textual description. The custom-trained model is obtained through a fine-tuning process where a pre-trained base model is trained further on curated data of digitalized artwork which was previously created by students. The fine-tuned model adjusts the weights of the base model so that it can now produce images in the artistic style of the peer group of students who contributed with their artwork. The workflow for image generation by the StableArtists app is presented in figure 4. The main goal of this system is to build AI literacy by helping students to acquire the knowledge necessary to understand AI from a technical, ethical and user or business needs perspective as described in UNESCO’s AI competency framework for students [34].

StableArtists allows the AI-based creation of artwork that reflects the diverse skills and styles of students from different age groups or backgrounds. The system is designed to be used in formal and non-formal educational settings. The user group consists of students under the guidance of a teacher or instructor. The development of the StableArtists system is motivated by an educational objective. Students shall learn to identify biases, acquire knowledge about ethical AI practices, and eventually become responsible citizens and remain independent actors in an increasingly AI-driven society. StableArtists provides the first use case to test the operability of the AI ethics assessment blueprint. We use the outcomes of the assessment for an ethics-by-design approach in the specification, technical realization of the diversity-sensitive AI system and its intended use. The following section is based on the results of the selected measures (c.f. appendix B) from the UNESCO ethical impact assessment  [33] following the processes of the AI ethics blueprint.

StableArtists has the dilemma to maximize two antagonistic metrics of the underlying AI model which are fairness and performance. Fairness is measured with respect to the representation of students who contribute artwork to the training data. Students are characterized by a set of features such as age, gender, ethnicity. The performance or quality of the model is measured by the mean esthetic value of the artwork composing the training data. For finetuned image-generation models we can fairly assume that the quality of the output is representative to the quality of the input training data. The quality, or synonymously the esthetic value, will be established by grading individual student’s contributions by (i) grading by the teacher, (ii) consensus decisions by the students, or (iii) by a multi-modal AI model acting as a “judge”. The students can vote on their preferred evaluation method. The finetuned model has the task to produce images of higher quality (mean grade) than a baseline model where all data would be included in the finetuning process. To fulfill this objective, artwork of lower grades must be removed from the training data which introduces selection bias. This exclusion bias will in turn lower the representativeness of the model with respect to students. The stakeholders of the system must agree on a bias mitigation strategy to select those images which will improve the esthetic value and still balance the representation of diverse student characteristics in the training data. Different bias mitigation strategies can be mapped by a materiality matrix assessment of the AI model as shown in figure 5. Students will understand how (cultural) bias may be inherent to generative AI systems as output bias is related to bias in the seen training data. StableArtists serves as a practical example through which students will gain insights into the ethical implications of AI technologies and developing a more responsible approach to their use.

While the Blueprint provides a promising foundation for AI ethics assessment, further research is needed to continue refining the framework. We propose three prospective directions:

1. 1.

   Develop a process for applying the blueprint in ethical assessment of potential transitions of AI applications between different risk categories with respect to the classification defined by the EU AI Act.

2. 2.

   Test the adaptability of the AI ethics assessment framework through use cases in: (i) different geographical zones with different interpretability of the ethics principles, and in (ii) sensitive areas like healthcare, recruiting, AI at the workplace, or collaborative AI systems.

3. 3.

   Identify generalisation aspects of AI Ethics Assessment across different application field sectors, with respect to the harmonization of outcomes, and guiding the standardisation of AI Ethics Assessment.