Implicit decision voting made by humans as normative and implementable rules with the help of language models

Hedfeld, Patrick Research Report Implicit decision voting made by humans as normative and implementable rules with the help of language models ifid Schriftenreihe: Beiträge zu IT-Management & Digitalisierung, No. 3 Provided in Cooperation with: ifid Institut für IT-Management & Digitalisierung, FOM Hochschule für Oekonomie & Management Suggested Citation: Hedfeld, Patrick (2025) : Implicit decision voting made by humans as normative and implementable rules with the help of language models, ifid Schriftenreihe: Beiträge zu ITManagement & Digitalisierung, No. 3, ISBN 978-3-89275-395-7, MA Akademie Verlagsund DruckGesellschaft mbH, Essen This Version is available at: https://hdl.handle.net/10419/315649 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0/ Implicit Decision Voting Made by Humans as Normative and Implementable Rules with the Help of Language Models ~ Patrick Hedfeld Rüdiger Buchkremer / Oliver Koch / Andreas Lischka (Hrsg.) Institut für IT-Management & Digitalisierung der FOM University of Applied Sciences ifid Schriftenreihe Beiträge zu IT-Management & Digitalisierung Band 3 Patrick Hedfeld Implicit Decision Voting Made by Humans as Normative and Implementable Rules with the Help of Language Models ifid Schriftenreihe der FOM, Band 3 Beiträge zu IT-Management & Digitalisierung Essen 2025 ISBN (Print) 978-3-89275-394-0 ISSN (Print) 2699-562X ISBN (eBook) 978-3-89275-395-7 ISSN (eBook) 2699-5638 Dieses Werk wird herausgegeben vom ifid Institut für IT-Management & Digitalisierung der FOM Hochschule für Oekonomie & Management gGmbH Verlag: MA Akademie Verlagsund Druck-Gesellschaft mbH, Leimkugelstraße 6, 45141 Essen [email protected] Die Deutsche Nationalbibliothek verzeichnet diese Publikation in der Deutschen Nationalbibliographie; detaillierte bibliographische Daten sind im Internet über http://dnb.d-nb.de abrufbar. Dieses Werk ist lizenziert unter CC BY 4.0: Creative Commons Namensnennung 4.0 International. Diese Lizenz erlaubt unter den Voraussetzungen der Lizenzbedingungen, u. A. der Namensnennung der Urheberin oder des Urhebers, der Angabe der CC-Lizenz (inkl. Link) und der ggf. vorgenommenen Änderungen die Bearbeitung, Vervielfältigung und Verbreitung des Materials in jedem Format oder Medium für beliebige Zwecke. Die Rechte und Pflichten in Zusammenhang mit der Lizenz ergeben sich ausschließlich aus dem Lizenzinhalt: CC BY 4.0 Deed | Namensnennung 4.0 International | Creative Commons | https://creativecommons.org/licenses/ by/4.0/legalcode.de. Die Bedingungen der Creative-Commons-Lizenz gelten nur für Originalmaterial. Die Wiederverwendung von Material aus anderen Quellen (gekennzeichnet mit Quellenangabe) wie z. B. von Schaubildern, Abbildungen, Fotos und Textauszügen erfordert ggf. weitere Nutzungsgenehmigungen durch den jeweiligen Rechteinhaber. Rüdiger Buchkremer / Oliver Koch / Andreas Lischka (Hrsg.) Implicit Decision Voting Made by Humans as Normative and Implementable Rules with the Help of Language Models Patrick Hedfeld Correspondence: E-Mail: [email protected] ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting II Foreword Artificial intelligence (AI) holds extraordinary potential to transform our world in profoundly positive ways. Whether in healthcare, education, or other critical domains, it provides innovative tools to tackle some of humanity’s most pressing challenges. I am deeply convinced that AI can and should be a force for good, but this vision hinges on one critical factor: the integration of ethics into every stage of its design and application. Ethics cannot be an afterthought; it is the bedrock upon which trust, fairness, and progress in AI systems must be built. This work delves into the fascinating interplay between implicit moral decisionmaking and AI. Specifically, it investigates how collective human decision votes can inform the development of ethical AI systems. By analyzing moral decision data, this research highlights the potential for AI to function as an implicit moral advisor—a system that honors human agency while fostering solutions that benefit all stakeholders. Through the application of generative language models, the study demonstrates how implicit moral preferences can be made transparent, shaped into normative principles, and positioned within broader societal dialogues. One of the most striking insights from this research is that even imperfect moral data, often shaped by human biases, can guide AI toward consensus-based ethical rules that advance societal well-being. By addressing critical challenges, such as algorithmic bias and the need for accountability, this work offers a framework for designing AI systems that function as ethical collaborators—partners in achieving shared human values, rather than neutral or purely utilitarian tools. As a professor and director of an AI institute, I view this work as an important contribution to an urgent conversation. It inspires us to develop AI systems that are not only technically innovative but also deeply aligned with humanity’s moral aspirations. Together, we can create AI technologies that empower individuals, strengthen communities, and act as a guiding force for ethical progress in our shared future. Essen, February 2025 Prof. Dr. Rüdiger Buchkremer Research Director of the FOM Institute for IT-Management and Digitalization (ifid) ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting III Abstract Can an ethically justifiable decision-making process be facilitated through a machine moral agent in the form of advice? The premise revolves around decision votes explicitly solicited from human individuals and based on scenarios such as the Trolley Problem, reflected in data and processed through generative language models. These advisories can then be formulated in a general manner and discussed within a societal context, emerging implicitly from individual decisions. Furthermore, we discuss the concept of an implicit moral agent and an honorable AI advisor. Keywords: AI, Generative Language Models, Decision Votes, Trolley Problem, Implicit Agent ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting IV Table of Contents Foreword .............................................................................................................. II Abstract ............................................................................................................... III List of Figures ...................................................................................................... V 1 Introduction ..................................................................................................... 1 2 Problem and Motivation .................................................................................. 2 2.1 Implicit Rules for Ethics........................................................................... 2 2.2 Racist Algorithms, the Importance and the Moral Side of Data .............. 4 3 Language Models in Machine Learning ......................................................... 6 3.1 From Simple Statistics to GPT Model ..................................................... 6 3.2 The Language Transformer Model as an Enabler .................................. 8 4 The Term Implicit .......................................................................................... 16 4.1 Implicit Data .......................................................................................... 16 4.2 Implicit Decisions and Decision Votes .................................................. 16 4.3 Implicit (Moral) Agents .......................................................................... 17 5 The Honorable AI Consultant ....................................................................... 19 5.1 Potential ................................................................................................ 19 5.2 Challenges and Limitations ................................................................... 22 5.3 Rules, Guidelines and Responsibility ................................................... 25 References ........................................................................................................ 27 ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting V List of Figures Figure 1: Mapping words in the moral data ...................................................... 11 Figure 2: First results, potentials and human acceptance ................................ 13 Figure 3: First Results (own representation) .................................................... 14 ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 1 1 Introduction In contrast to racist texts, moral data often contain implicit decision voting made by humans. This can be demonstrated by general language models, which leverage brain-inspired and generative approaches. The results produced by these models have the potential to be accepted by humans as normative and implementable rules. Language models, acting as implicit moral advisors, utilize moral data derived from win-lose decisions to address issues. By making the results of implicit voting visible, they pave the way for win-win decisions and situations. However, it is important to engage in thorough discussions because moral data created by humans encompass various factors, including feelings, erroneous decisions, and egoism. Despite this complexity, a theoretical implicit moral advisor has the capacity to lead to normative and implementable rules through careful consideration and discussion. ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 8 word you enter and pay attention. We would call this context of the information e.g. the sentence: I am a student will be translated to French: Je suis étudiant. This means that I am a maps to je suis without considering two or three words but giving attention to the whole struct. In short, the attention algorithms are introducing queries, keys and values for every word (Vaswani et al., 2017: 3). With these three new parameters it is possible to give more or less attention to every word. As a last step transformers were established and they used attention algorithms including self-attention algorithms. The idea of self-attention is to figure out how the sentence is mapped to itself e.g. the word thinking might have a bigger score on the I than on all the other words in the same sentence. Or in other words: In the sentence I was thinking of a new house. – you will have scores on every connection and relation between the words like: Who was thinking? What was it? and so on (Vaswani et al., 2017: 6). As a last step the general pre-trained transformer were established. We would like it to use it in the investigation (cf. Radford / Narasimhan, 2018a). The more parameters the model has the more language it can produce (cf. Tunstall et al., 2022). We will use GPT-2 which has 1.5 billion parameters in the network for use which can be used for implicit language generation (cf. Huggingface, 2022; cf. Radford et al., 2018b). The API has still 124 million parameters based on a very large corpus of English data (cf. Huggingface, 2022). The number of parameters, the volume of data, and the linguistic reference are entirely adequate for the objectives of this paper. In future deliberations, it may be possible to opt for more advanced models or networks with increased parameters and linguistic input. You can use the GPT2 model transformer as interface to humans because it generates text which can be read by humans and it learns very fast even compared to other language approaches (cf. Kojima et al., 2022). 3.2 The Language Transformer Model as an Enabler Based on the ideas of G.W.F. Hegel, one of the most crucial human abilities is conceptual work and its influence on our thinking and social interactions. In Hegel’s philosophy, freedom holds a unique significance, as it emerges through the evolution of the spirit in more advanced forms of thought and action (cf. Seeberger, 1961). In systems philosophy, the term psychology appears in the subjective geist and is established there by intelligence and will, which unfold on ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 9 ever new conceptual levels (Drüe et al., 2000: 274-283). In the will gradually unfold or developed Genuss (pleasure), Neigung (tendency), freiere Aktivitäten (free activities), which reach into the objektiver Geist (objective geist) where one finds, among other things, the legal concept in the science of the human beings Philosophie des Zwischenmenschlichen (cf. Jaeschke, 2010: 363). This method of self-movement of the term is also discussed in the literature (cf. Röttges, 1976; Drüe, 2000: 283). The philosophy of reflection of the german idealism now assumes that on the one hand it is about the Sollen (should) but on the other hand also about the implementation (Lütge, 2002: 243; Hegel, 1986 [1801]: §68) because the philosophy should not stop at the pure term (cf. Hegel, 1986 [1821a]; 1986 [1821b]) because philosophy has to do with the idea which is not powerless in order to only ought and not to really be (cf. Hegel, 1986 [1830] §6). Hegel’s philosophy demonstrates the dialectical process that is crucial for connecting individuals within the subjective spirit (subjektiver Geist) as they transition to institutionalized forms within the objective spirit (objektiver Geist) and other societal structures. This influenced, among other things, Homann’s ideas on business ethics. In the concept of a conversion paradigm, as proposed by Homann (Homann, 2002: 189), the emphasis is on individual actions. It revolves around the idea of the canon of duties and virtues, aiming to overcome tendencies to act solely out of duty by providing arguments and good reasons, even in cases of weak motivation to act. This involves three main aspects: firstly, recognizing (or rejecting) morality; secondly, defining actions communicatively; and thirdly, motivating oneself through informal coercion to adhere to these actions (cf. Homann, 2002: 190). A morally data-driven consultant can now do various things: propose normative ideas and principles, which are based on the implicit decision votes of the individually made decisions in the trolley problem e.g. people are preferable to animals or the larger group should usually be spared. If I’m not in a specific situation but have some time before making a decision, then these general guidelines can be unanimously accepted by everyone. This is because, on one hand, we all share a common humanity, and on the other hand, there’s a high likelihood that a larger group of people would agree with these principles (e.g. in the simulations of consensus Homann/Lütge 2013: 83). The problem on the moral data or the implicit decisions might be that everything should be occurrences within the data: drives, inclinations, wrong decisions, in the idea on the development of the different stages etc. The moral computer-based consultant can also be seen as ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 10 a novel tool that addresses existing limitations (cf. Homann, 2003) and opens up new avenues of thought. This is because the additional resources provided by the computer can now be integrated into the decision-making process, expanding the possibilities for consideration. Implicit decision-making votes can be made visible in extreme situations and the action is not always justifiable for the acting actor because of its implicit structure (Minnameier 2016: 79). The idea of giving incentives is like an iceberg floating under the surface of the water (Homann, 2014: 232; cf. Minnameier, 2005). However, this should serve a free decision-making (Homann, 2014: 14) and not be nudgy (cf. Thaler / Sunstein, 2009). It should serve about a connection between empiricism (the decision votes prepared by the moral data-driven consultant) and the normative values, which should lead to normative judgements (Normative Urteile) (Homann, 2014: 14). In addition to Handlungsethik (action ethics) and Ordnungsethik (normative ethics), there is also the possibility of a third level, which can be understood as a human level, which in turn is available as a discourse about the norms or rules of the game (Pies, 2009: 11). The moral advisor could display decision votes or even offer default settings in certain situations, with the understanding that individuals have the opportunity to justify their decisions against these defaults for valid reasons. One goal could be to foster moral character development through habituation or the cultivation of virtues. This development is facilitated by the visible decision votes provided by the moral advisor (Homann, 2014: 242), without rigidly enforcing norms or declaring them universally valid indefinitely. ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 11 Figure 1: Mapping words in the moral data Source: Illustration Based on (Vig / Belinkov, 2019). The data provided by moral machines (cf. Awad, 2018b) can be used to establish a moral data-driven level 2 implicit consultant (or agent) according to Moor (cf. Moor, 2006). The advantages of this approach are manifold: Firstly, the advisor operates implicitly, allowing individuals the freedom to reject or modify its suggestions as desired (Homann, 2014: 214). Secondly, the data provided is based on decision votes that may initially appear as win-lose situations, but can ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 12 be evolved into win-win outcomes through generalization. Thirdly, the language model is inspired by the human brain, enabling it to generate both simple judgments or decisions, as well as guidelines and general statements. Ultimately, human acceptance is crucial to ensure consensus and acceptance of the corresponding rules among people. The moral consultant represents a solution to scarcity by leveraging computer-controlled performance (cf. Homann, 2003). The structure that now follows is based on the GPT (See Moral Machines file: MMdataReadMe.txt for a detailed description at Bonnefon 2018) API and is trained using the data from the trolley problem. The data provide different scenarios, these are: Utilitarian, Gender, Fitness, Age, Social Value, Species and Random. Utilitarian represents a more or less problem, gender a different gender situation (male and female situations), fitness a difference between man and athlete man for example, age represents situations of different ages, for example old woman and woman, social value includes, among other things, comparisons with the an executive e.g., species compares humans and animals with each other (also in combination) and random was a remnant of a first data collection and contains several variations. The data for the set random is not used for training the data and the whole data is divided into three groups: First: training data, Second: valuation data and Third: test data. The data is first processed and the answers are brought from two lines (the work’s storage method) onto one line and is sorted. In a second step, the very structured data is translated into text. An advantage of language models is usually unstructured data, in this case the data is highly structured and prepared. A pre-trained GPT2 model is loaded via the API and then trained with initially ten thousand moral decisions in three epochs. For the test and the validation data, a few variants with two and three people in the difference are taken out of the utilitarian scenarios. To give an example of the mapping: (data fields: Saved = 1, Intervention = 0) One men, two women and two dogs in the data will be in this in text: If the car stays one men, two women and a dog are saved. In order to test the model, the validation and the accuracy are measured and a mathematical confusion matrix is generated (a confusions matrix tests the different states for correctness, in this case it is a 2x2 matrix). In the first field, test data are used to test the first state in the second field (stay on stay) then cross (swerve on stay). Due to the highly structured nature of the data, the accuracy values approach one hundred percent and the losses are less ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 13 than one percent. Additionally, the confusion matrix closely resembles the identity matrix (You might think of a matrix with four fields in it. The first field represents all stay actions regarding the stay actions and so on. This means stay = stay is very high and swerve = swerve possibly but stay = swerve is very low in both cases). Figure 2: First results, potentials and human acceptance The initial results were divided into three groups. Firstly, logical statements (1) were included in the moral data to verify if the model is functioning correctly. The expectation was for the majority of statements in the trained model to reproduce these logical statements accurately. Secondly, logical tests (2) and (3) aimed to assess if the model could generalize effectively. These tests involved scenarios not present in the original data, such as a pregnant man or a group larger than five individuals. The results showed a clear trend consistent with moral decision- ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 14 making principles, including a preference for saving more people, as observed in evaluations by moral machines. This suggests that the model successfully reinforces logical connections during training. In simpler terms, if there’s a connection between a pregnant woman and a woman, the model also establishes a stronger connection between a woman and a man, thereby implying a connection between a pregnant woman and a man. Thirdly (4), (5), the model was asked about general norms. Since there must be a connection between man, woman etc. and human or humanity (or a connection to dog and cat to animal) and the fact of more or less is also shown, the two statements come about. In this way, a utilitarian approach is basically preserved but on a generic level regarding Buchanan as choice within the rules/choice of rules (cf. Buchanan, 1984). Figure 3: First Results (own representation) It provides valuable data for refining the moral advisor. In addition, other language models (e.g. BERT) can also be used to check sentences for their implicit acceptance by the model (cf. Devlin et al., 2018). These sentences can also be examined humanely. This approach may look similar to training a harmless counselor, but it serves the purpose of finding or discussing general rules (cf. Bai et al., 2022) in respect of the idea of RLHF (Reinforcement Learning with Human Feedback). The concept of making probabilities visible also enhances moral education and further development, as it introduces the decision votes of individuals into the discussion, making them impossible to ignore. This transparency encourages engagement and fosters deeper understanding of the moral implications of decisions. Making it impossible to study a technology without the value-system of the community (cf. Martin / Freeman, 2004) maybe in a triangle of business, ethics and technology. The idea is that our quick decision system the so-called ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 15 system 1 is more at risk of having a racist bias than the slow system 2 (cf. Kahneman, 2011; Agan et al., 2023). For this reason, education can also serve to use a moral, data-driven advisor or bring it at least into the discussion based on human acceptance, freedom and even a strong system 2 (slow system). ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 16 4 The Term Implicit The term implicit in this paper occurs three times. First: implicit data. Second: implicit decisions and decision votes. Third: implicit agents. These terms must be kept apart. I would like to start with the concept of implicit knowledge based on the thoughts of Nonaka and Takeuchi (cf. Nonaka / Takeuchi, 2012). Implicit knowledge here is an unused resource or something that can be made visible or explicit; it is described as premonition, intuition or a mental or cognitive understanding of the world (ibid.: 24). For the sake of simplicity, you can also write: Implicit is everything that does not appear explicitly. 4.1 Implicit Data Modern neural networks require data for training. This paper addresses the question: What constitutes moral data? It is relatively straightforward to identify what constitutes racist data – examples abound, such as discrimination based on skin color, height, religion, or gender. But what defines moral data, essential for making morally sound, or at least better, decisions? (cf. Floridi / Taddeo, 2016). Furthermore, data can reveal implicit elements that may not be immediately apparent. My movement data reveals more about me than I might want to reveal. From Monday to Friday, I follow a routine of going to work and almost always sleeping at home in the evenings. Occasionally, I might meet my friends every Saturday at a sports stadium, where you might even discern my favorite club and more. Over time, observing me reveals implicit aspects in data that can offer insights into me or society beyond what we consciously realize. This understanding can increase the likelihood of accepting rules derived from our own data, as they reflect our behaviors and preferences implicitly. 4.2 Implicit Decisions and Decision Votes For many years, researchers have studied whether elections could be predicted by better analyzing undecided voters (cf. Lundberg / Payne, 2014). Attitudes, ideas, or unconscious actions can provide valuable insights into preferences, including the choice of a political party or candidate (cf. Friese et al., 2012). Implicit decisions made by individuals have the potential to evolve into societal norms or rules that are accepted with consensus. The focus shifts from “What would I do ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 17 in a situation?“ to “What would be beneficial for a person if this were to occur in this situation?“. In this context, decision data or data from individuals can serve as the basis for statistical studies or large language models. The resulting rules can then be subject to social discussion or evaluation of their ability to garner consensus. Ultimately, the establishment of these rules or norms relies on free human acceptance. 4.3 Implicit (Moral) Agents According to Moor, there exist different moral agents, among which the implicit ethical agent is worth considering. Moor suggests that if one wishes to instill ethics into a machine, one way is to constrain the machine’s actions to prevent unethical outcomes. In this approach, machine ethics is achieved by creating software that implicitly promotes ethical behavior, rather than explicitly containing ethical maxims. The machine behaves ethically because its internal functions naturally lead to ethical behavior or, at the very least, prevent unethical behavior. Ethical behavior becomes inherent to the machine’s nature, embodying virtues to some extent. Computers can serve as implicit ethical agents when their design prioritizes safety or critical reliability concerns. For instance, automated teller machines and web banking software act as agents for banks, performing many tasks of human tellers and sometimes more. Given the ethical significance of transactions involving money, these systems must adhere to ethical standards. Machines must be carefully constructed to give out or transfer the correct amount of money every time a banking transaction occurs. A line of code telling the computer to be honest won’t accomplish this. Aristotle suggested that humans could obtain virtue by developing habits. But with machines, we can build in the behavior without the need for a learning curve. Of course, such machine virtues are task specific and rather limited. (Cohn et al., 2022; Moor, 2006: 19) Indeed, in 2006, Moor could not have foreseen major language models or their potential benefits. It can be argued that modern chat software exists somewhere between the implicit and explicit agent, as it engages explicitly with human interaction. Moor outlines four levels: the ethical impact agent, implicit ethical agent, explicit ethical agent, and full ethical agent. It is important to note that the implicit ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 24 programs or decisions regarding credit limits (Golbin / Rao, 2019; Kordzadeh / Ghasemaghaei, 2022:1-11). Another challenge in developing the honorable AI consultant is the evolution of morality over time. What is currently considered to conform to ethical standards may be questioned and deemed morally problematic in the future. Therefore, continuous monitoring and updating of the system’s moral values are necessary to support moral development and progress in a measured manner, ensuring that the AI reflects evolving ethical perspectives (cf. Lara / Deckers, 2020: 279). Another difficulty arises from the chosen method of moral implementation. In experiments such as the Moral Machine, which served as inspiration, it became evident that even seemingly clear preferences exhibited cultural differences. (Awad et al., 2018a: 59-63). Given that companies often operate within multicultural environments or have the potential to do so, diversification of ethical perspectives may arise in this context. Differences in ethical stances can be expected. However, predicting such diversification is challenging, as the moral decisions made in experiments like the Moral Machine differ fundamentally from those applicable to the honorable AI consultant. In the proposed concept, decisions would pertain to moral dilemmas within an economic context rather than decisions concerning life-threatening situations. Therefore, this presents a potential challenge that may or may not materialize. It is worth mentioning worth that Awad et al. (2018a: 59-63) have found that large parts of the world show some agreement in their ethical preferences. It is important to understand how the honorable AI consultant affects human decisions. There’s a danger that people might get hurt if they rely too much on its recommendations without questioning or thinking carefully about them (cf. Busuioc, 2021: 26). This excessive dependence on AI system recommendations is known as automation bias—the tendency to uncritically trust automated decisions over human judgment. This phenomenon occurs when users overly trust the decisions made by automated support systems, leading to decreased vigilance in seeking and processing information (Busuioc, 2021: 26; Lyell / Coiera, 2017: 423). Krügel / Ostermaier / Uhl, 2022: 1-3) also found that users trust the ethical advice of an AI even without information about the training data. Interestingly, this trust remains even when users have information that could potentially raise doubts about the system’s reliability. Their study suggests that people are more likely to place excessive trust in an AI than to distrust it. (Krügel et al., 2022:1-20). ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 25 5.3 Rules, Guidelines and Responsibility The preceding section highlighted the challenges and constraints inherent in the development and utilization of an honorable AI consultant, as well as those encountered with AI systems in general. AI is becoming increasingly integrated into various aspects of human life, bringing about significant transformations with profound impacts on numerous social domains. The importance of cultivating ethically guided programming and implementation of such systems is becoming increasingly evident. In recent years, various educational institutions, private companies, and public sector organizations have developed and published guidelines for ethical AI. Jobin, Ienca and Vayena (Jobin / Ienca / Vayena, 2019: 389) came to the conclusion that there is global agreement on five ethical fundamental principles. These are (1) transparency, (2) fairness and justice, (3) non-harm, (4) responsibility and (5) data protection (cf. Jobin et al., 2019: 389). In their study, they emphasize the substantial divergence in interpretations of these principles, their applicability to different topics, domains, or stakeholders, the appropriate methods for implementation, and why adherence to these principles is deemed important (cf. Jobin et al., 2019: 389). To ensure adherence to the five principles despite differences, we can refer to Sarah Spiekerman’s explanation (2021: 248). A crucial aspect is the requirement for transparency in the functionality and decision-making processes of artificial systems. This necessitates documentation and communication about their operations and decision-making to be appropriate, accessible, comprehensive, meaningful, and truthful (cf. Spiekermann, 2016: 59; cf. Spiekermann, 2019). Ensuring equal treatment of all users and avoiding systematic bias through AI are central requirements within the framework of ethical principles such as fairness and justice. It is also vital to ensure equal rights and accessibility of the system for all users. The ethical principle of non-harm is particularly relevant to system security, emphasizing the importance of preventing any damage or negative effects on users or society. Responsibility entails that those overseeing AI systems are accountable and responsible for their actions. Integrity is crucial to ensure that responsible actions are carried out ethically and with integrity. Finally, the ethical principle of data stewardship pertains to data handling, encompassing principles such as those outlined in the European General Data Protection Regulation, including the right to be forgotten, data portability, informed access, among others (cf. Spiekermann, 2021: 248). Ethical guidelines for AI offer a framework to strike a balance between harnessing the diverse capabilities of AI and ensuring oversight over its development and ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 26 impact. The social acceptance of AI technologies hinges on whether the benefits are deemed significant and whether the risks are perceived as avoidable, reducible, or controllable. (cf. Floridi et al., 2018: 694). Through the introduction and implementation principles and thus the creation of an ethical AI – also responsible called conscious AI – the hope is to build trust in the systems as well to limit negative effects (cf. Eitel-Porter, 2021: 73). It is also important meaning to take responsibility through such guidelines, which type of AI is developed and how it is used (cf. Floridi et al., 2018: 692). In the future there might be a need for standardization of such ethical guidelines to ensure the safe use of AI (cf. Jobin et al., 2019: 396-397). In addition to ethical guidelines, there is a strong emphasis on enhancing digital skills and promoting the responsible use of algorithms (cf. Krügel et al., 2022: 21). The implementation of education initiatives for users could be considered as a possible solution (cf. Burrell, 2016: 10) to have robust governance and compliance mechanisms in place to integrate corporate structures to avoid unwanted negative effects of AI systems (cf. Eitel-Porter, 2021: 73). A typical recommendation for one appropriate leadership structure to ensure responsible AI is to establish a two-tier structure at the top. On one hand, it is recommended to establish an ethics council or advisory group to incorporate external contributions from society. On the other hand, it is proposed to establish an ethics committee or review board internally to guide and monitor the focus on responsible AI (cf. de Laat , 2021: 163). Companies also need support by employees who continually work to ensure that AI systems are operate in a measured, safe and responsible manner (cf. Wilson / Daugherty, 2018: 11). There are concrete steps to implement a robust leadership structure for example Ray EitelPorter (cf. 2021: 73-80) in his article Beyond the promise: implementing ethical AI. In conclusion, adhering to ethical guidelines throughout the development and utilization stages of the honorable AI consultant, alongside the implementation of suitable leadership and control structures, can assist in mitigating the challenges that arise and ensure responsible usage of the system. An initial step could involve integrating an implicit moral agent using data from human decision-making, which could implicitly align with societal rules and norms, leading to a win-win situation. Subsequently, the introduction of an honorable AI consultant could further contribute to the establishment of acceptable and implementable rules for society. ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 27 References Agan, A. Y. / Davenport, D. / Ludwig, J. / Mullainathan, S. (2023): Automating Automaticity: How the Context of Human Choice Affects the Extent of Algorithmic Bias, National Bureau of Economic Research, Cambridge, No. w30981. Aharoni, Y. / Tihanyi, L. / Connelly, B. L. (2011): Managerial decision-making in international business: A forty-five-year retrospective. Journal of World Business, 46(2), 135–142. https://doi.org/10.1016/j.jwb.2010.05.001. Ananny, M. (2016): Toward an ethics of algorithms: Convening, observation, probability, and timeliness. In: Science, Technology, & Human Values, 41(1), 93-117. Artinger, F. / Petersen, M. / Gigerenzer, G. / Weibler, J. (2015): Heuristics as adaptive decision strategies in management. Journal of Organizational Behavior, 36 (S1), 33-52. https://doi.org/10.1002/job.1950. Awad, E. / Dsouza, S. / Kim, R. / Schulz J. / Henrich, J. / Shariff, A. / Bonnefon, J.F. / Rahwan, I. (2018a): The Moral Machine experiment. In: Nature 563, 59–64. https://doi.org/10.1038/s41586-018-0637-6. Awad, E. (2018b): Moral Machine. https://osf.io/3hvt2/. Bai, Y. / Jones, A. / Ndousse, K. / Askell, A. / Chen, A. / DasSarma, N. / Kaplan, J. (2022): Training a helpful and harmless Assistant with reinforcement Learning from human Feedback. arXiv preprint arXiv: 2204.05862. Bartneck, C. / Lütge, C. / Wagner, A. / Welsh, S. (2019): Ethik in KI und Robotik. München: Carl Hanser Verlag. Bajohr, H. (2022): Schreibenlassen. Texte zur Literatur des Digitalen. Berlin: August Verlag. Bengio Y. / Simard, P. / Frasconi, P. (1994): Learning long-term dependencies with gradient descent is difficult. In: IEEE Transactions on Neural Networks Vol. 5 (2), 157-166, March 1994, https://ieeexplore.ieee.org/document/279181. Bentham, J. / Mill, J.S.(2004): Utilitarianism and other essays. London: Penguin UK. ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 28 Bonnefon, J.-F. / Shariff, A. / Rahwan, I. (2016): The social dilemma of autonomous vehicles. In: Science, 352(6293), 1573–1576. https://doi.org/10.1126/science.aaf2654. Bowker, G. C. / Star, S. L. (2000): Sorting things out: Classification and its consequences. MIT press. Buchanan, J.M. (1984): Die Grenzen der Freiheit: Zwischen Anarchie und Leviathan, Tübingen: Mohr. Burrell, J. (2016): How the machine ‘thinks’: Understanding opacity in machine learning algorithms. Big Data & Society, 3(1), 2053951715622512. https://doi.org/10.1177/2053951715622512. Bushman, B. J. / DeWall, C. N. / Pond, R. S. / Hanus, M. D. (2014): Low glucose relates to greater aggression in married couples. Proceedings of the National Academy of Sciences, 111(17), 6254–6257. https://doi.org/10.1073/pnas.1400619111. Busuioc, M. (2021): Accountable Artificial Intelligence: Holding Algorithms to Account. Public Administration Review, 81(5), 825–836. https://doi.org/10.1111/puar.13293. Cohn, A. / Gesche, T. / Maréchal, M. A. (2022): Honesty in the digital age. In: Management Science 68(2), 827-845. Danziger, S. / Levav, J. / Avnaim-Pesso, L. (2011): Extraneous factors in judicial decisions. Proceedings of the National Academy of Sciences, 108(17), 6889 6892. https://doi.org/10.1073/pnas.1018033108. Davis, M.(1958): Computability and Unsolvability, New York: McGraw-Hill. de Laat, P. B. (2021). Companies Committed to Responsible AI: From Principles towards Implementation and Regulation? Philosophy & Technology, 34(4), 1135–1193. https://doi.org/10.1007/s13347-021-00474-3. Deckert, R. / Meyer, E. (2020): Digitalisierung und Künstliche Intelligenz: Kooperation von Menschen und Maschinen aktiv gestalten. Springer Fachmedien. https://doi.org/10.1007/978-3-658-31795-9. Del Rosario, M.B. / Redmond, S.J. / Lovell, N.H. (2015): Tracking the Evolution of Smartphone Sensing for Monitoring Human Movement. In: Sensors, 15, 18901-18933, Österreichische Artificial-Intelligence-Tagung (KONNAI) Salzburg, Österreich, 18.–21. September 1990 Proceedings, 252, 191. https://doi.org/10.3390/s150818901. ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 29 Devlin, J. /Chang, M.W. /Lee, K. /Toutanova, K. (2018): Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. Drüe H. / Gethmann-Siefert A. / Hackenesch, C. / Jaeschke, W. / Neuser W. / Schnädelbach H. (2000): Hegels Enzyklopädie der philosophischen Wissenschaften (1830), Frankfurt am Main: Suhrkamp. Eitel-Porter, R. (2021): Beyond the promise: Implementing ethical AI. AI and Ethics, 1(1), 73–80. https://doi.org/10.1007/s43681-020-00011-6. Ferrell, O. C. / Fraedrich, J. / Ferrell, L. (2021): Business Ethics: Ethical Decision Making and Cases. Cengage Learning. Friedman, B. / Nissenbaum, H. (1996): Bias in computer systems. ACM Transactions on Information Systems, 14(3), 330–347. https://doi.org/10.1145/230538.230561. Friese, M. / Smith, C. T. / Plischke, T. / Bluemke, M., / Nosek, B. A. (2012): Do implicit attitudes predict actual voting behavior particularly for undecided voters?. PLoS ONE 7(8): e44130. https://doi.org/10.1371/journal.pone.0044130 Floridi, L. / Taddeo, M. (2016): What is data ethics?. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 374(2083), 20160360. Floridi, L. / Cowls, J. / Beltrametti, M. / Chatila, R. / Chazerand, P. / Dignum, V. / Luetge, C. / Madelin, R. / Pagallo, U. / Rossi, F. / Schafer, B. / Valcke, P. / Vayena, E. (2018): AI4 People—An Ethical Framework for a Good AI Society: Opportunities, Risks, Principles, and Recommendations. Minds and Machines, 28(4), 689–707. https://doi.org/10.1007/s11023-018-9482-5. Foot, P. (1978): The Problem of Abortion and the Doctrine of the Double Effect, In: Philippa Foot: Virtues and Vices and Other Essays in Moral Philosophy, Oxford: Oxford University Press, 19–32. Gigerenzer, G. / Brighton, H. (2009): Homo Heuristicus: Why Biased Minds Make Better Inferences. Topics in Cognitive Science, 1(1), 107–143. https://doi.org/10.1111/j.1756-8765.2008.01006.x. Gigerenzer, G. / Gaissmaier, W. (2011): Heuristic Decision Making. Annual Review of Psychology, 62(1), 451–482. https://doi.org/10.1146/annurev-psych120709-145346. ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 30 Giubilini, A. / Savulescu, J. (2018): The Artificial Moral Advisor. The “Ideal Observer” Meets Artificial Intelligence. Philosophy & Technology, 31(2), 169– 188. https://doi.org/10.1007/s13347-017-0285-z Gips, J.(1994). Toward the ethical robot. https://philarchive.org/rec/GIPTTE. Golbin, I. / Rao, A. (2019): What is fair when it comes to AI bias? Strategy+business. https://www.strategy-business.com/article/What-is-fair-when-itcomes-to-AI-bias. Haque T. U. / Saber N. N. / Shah F. M.(2018): Sentiment analysis on large scale Amazon product reviews. IEEE International Conference on Innovative Research and Development (ICIRD) Bangkok, Thailand, 1-6, https://ieeexplore.ieee.org/document/8376299. Hedfeld, P. (2019): Die neuen Glasperlen. Neue Gesellschaft Frankfurter Hefte, Jg. 2019, Nr. 10, 30-32, Berlin: Dietz Verlag. Hegel, G.W.F. (1801, 1986): Jenaer Schriften 1801-1807 Berlin: Suhrkamp Verlag. Hegel, G.W.F. (1821a, 1986): Grundlinien der Philosophie des Rechts Berlin: Suhrkamp Verlag. Hegel, G.W.F. (1821b, 1986): Vorlesungen über die Geschichte der Philosophie III Berlin: Suhrkamp Verlag. Hegel, G.W.F. (1830, 1986): Enzyklopädie der philosophischen Wissenschaften I Berlin: Suhrkamp Verlag. Hochreiter, S. / Schmidhuber, J. (1997): Long Short-term Memory. In: Neural computation. 9. p.1735-1780. 10.1162/neco.1997.9.8.1735. Homann, K. / Lütge, C. (2003): Anreize und Moral. Gesellschaftstheorie–EthikAnwendungen. Philosophie und Ökonomik Band 1. Münster: LIT Verlag. Homann, K. / Lütge C. (2013): Einführung in die Wirtschaftsethik (3rd. Ed.) Berlin: LIT Verlag Dr. W. Hopf. Homann, K. (2014): Sollen und Können Wien: Ibera Verlag. Homann, K. (2020): Praktische Philosophie und ökonomische Theorie – Aufsätze und Vorträge Berlin: LIT Verlag Dr. W. Hopf. Hopfield, John. (1982): Neural Networks and Physical Systems with Emergent Collective Computational Abilities. In: Proceedings of the National Academy ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 31 of Sciences of the United States of America. Vol. 79. p. 2554-2558. 10.1073/pnas.79.8.2554. HuggingFace (2022): GPT2. https://huggingface.co/gpt2. Jaeschke, W. (2016): HEGEL-Handbuch: Leben–Werk–Schule (2nd. edition). Stuttgart: Springer-Verlag. Jobin, A. / Ienca, M. / Vayena, E. (2019): The global landscape of AI ethics guidelines. Nature Machine Intelligence, 1(9), 389-399. https://doi.org/10.1038/s42256-019-0088-2. Jordan, M I. (1986): Serial order: a parallel distributed processing Approach. Technical report, June 1985-March 1986. United States: N. p. Kahneman, D. (2011): Thinking, fast and slow. Macmillan. Kant, I. (1870): Grundlegung zur Metaphysik der Sitten (Vol. 28). L. Heimann. Kaffka, T. (2017): Neuronale Netze Grundlagen. mitp: Frechen. Kordzadeh, N. / Ghasemaghaei, M. (2022): Algorithmic bias: Review, synthesis, and future research directions. European Journal of Information Systems, 31(3), 388–409. https://doi.org/10.1080/0960085X.2021.1927212 . Kouchaki, M. / Desai, S. D. (2015): Anxious, threatened, and also unethical: How anxiety makes individuals feel threatened and commit unethical acts. Journal of Applied Psychology, 100(2), 360–375. https://doi.org/10.1037/a0037796 . Kojima, T. / Gu, S. / S., Reid M. / Matsuo, Y. / Iwasawa, Y.): Large language models are zero-shot reasoners. https://doi.org/10.48550/arXiv.2205.11916. Krügel, S. / Ostermaier, A. / Uhl, M. (2022): Zombies in the Loop? Humans Trust Untrustworthy AI-Advisors for Ethical Decisions. Philosophy & Technology, 35(1), 17. https://doi.org/10.1007/s13347-022-00511-9 . Lara, F. / Deckers, J. (2020): Artificial Intelligence as a Socratic Assistant for Moral Enhancement. Neuroethics, 13(3), 275–287. https://doi.org/10.1007/s12152-019-09401-y . Lepri, B. / Oliver, N. / Pentland, A. (2021): Ethical machines: The human-centric use of artificial Science. 2021 Mar 3;24(3):102249. doi: 10.1016/j.isci.2021.102249. PMID: 33763636; PMCID: PMC7973859 url: https://pubmed.ncbi.nlm.nih.gov/33763636/ ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 32 Lundberg, K. B., / Payne, B. K. (2014): Decisions among the undecided: Implicit attitudes predict future voting behavior of undecided voters. PloS one, 9(1), e85680. Lütge, C. (Ed.) /Karl Homann (2002): Vorteile und Anreize. Tübingen: Mohr Siebeck. Lutz, T. (1959): Stochastische Texte. https://auer.netzliteratur.net/0_lutz/lutz_original.html. Lyell, D. / Coiera, E. (2017): Automation bias and verification complexity: A systematic review. Journal of the American Medical Informatics Association, 24(2), 423–431. https://doi.org/10.1093/jamia/ocw105. Martin, K. E. / Freeman, R. E. (2004): The separation of technology and ethics in business ethics. Journal of Business Ethics, 53, 353-364. Minnameier, G. (2005): Wer Moral hat, hat die Qual, aber letztlich keine Wahl!: Homanns (Wirtschafts) Ethik im Kontext der Wirtschaftsdidaktik. Zeitschrift fürs Berufsund Wirtschaftspädagogik 101, 19-42. Minnameier, G. (2016): Moralische Motivation und ökonomische Rationalität. In: Ethik und Beruf: Interdisziplinäre Zugänge, 73. Bielefeld: wbv Media GmbH & Company KG. Misselhorn, C. (2018): Grundfragen der Maschinenethik. Ditzingen: Reclam. Misselhorn, C. (2021): Künstliche Intelligenz und Empathie. Ditzingen: Reclam. Mitchell T. (1997): Machine Learning. New York: McGraw-Hill Education International Edition. Mittelstadt, B. D. / Allo, P. / Taddeo, M. / Wachter, S. / Floridi, L. (2016): The ethics of algorithms: Mapping the debate. Big Data & Society, 3(2), 2053951716679679. Moor, J.H. (2006): The Nature, Importance, and Difficulty of Machine Ethics. Machine Ethics IEEE, August 2006, 18-21. Nonaka, I. / Takeuchi, H. (2012): Die Organisation des Wissens, wie japanische Unternehmen eine brachliegende Ressource nutzbar machen. Frankfurt am Main: Campus Verlag. Oinkina / Hakyll (2015): Understanding LSTM Networks. https://colah.github.io/posts/2015-08-Understanding-LSTMs/. ifid Schriftenreihe, Bd. 3, Hedfeld: Implicit Decision Voting 33 Pies, I. (2009): Moral als Heuristik. Ordonomische Schriften zur Wirtschaftsethik Berlin: WVB, Wissenschaftlicher Verlag. Powers, T. M.(2006): Prospects for a kantian machine. Intelligent Systems, IEEE Volume 21(4), August 2006, 46-51. Radford, A. / Narasimhan, K.(2018a): Improving Language Understanding by Generative Pre-Training. Radford, A. / Wu, J. / Child, R. / Luan, D. / Amodei, D. / Sutskever, I. (2018b): Language Models are Unsupervised Multitask Learners. Ramge, T. (2018): Mensch und Maschine: Wie Künstliche Intelligenz und Roboter unser Leben verändern. Ditzingen: Reclam. Ransbotham, S. / Candelon, F. / Kiron, D. / LaFountain, B. / Khodabandeh, S. (2021): The cultural benefits of artificial intelligence in the Enterprise. MIT Sloan Management Review and Boston Consulting Group: Cambridge, MA, USA. Rayburn, W. M., / Diederich, J. (1990): Some Remarks on Emotion, Cognition, and Connectionist Systems. Konnektionismus in Artificial Intelligence und Kognitionsforschung: 6.Österreichische Artificial-Intelligence-Tagung (KONNAI) Salzburg, Österreich, 18.–21. September 1990 Proceedings, 191-195. Berlin & Heidelberg : Springer Verlag. Rojas, R. (1993): Theorie der neuronalen Netze. Eine systematische Einführung. Berlin: Springer-Lehrbuch. Röttges, H. (1976): Der Begriff der Methode in der Philosophie HEGELs. Meisenheim am Glan: Verlag Anton Hain. Riek, L. / Howard, D. (2014): A Code of Ethics for the Human-Robot Interaction Profession. Proceedings of We Robot 2014. https://ssrn.com/abstract=2757805. Rey, G. / Wender, K. (2011): Neuronale Netze. (2nd Ed.) Bern: Huber. Rumelhart, D. / Hinton, G. / Williams, R. (1985): Learning internal representations by error propagation. Tech. rep. ICS 8504. San Diego, California: Institute for Cognitive Science, University of California. Sandvig, C. / Hamilton, K. / Karahalios, K. / Langbort, C. (2016): Automation, Algorithms, and Politics when the Algorithm itself is a racist: Diagnosing ethical