Ethics and Law for AI PDF

ETHICS, LAW AND AI Prof. Redaelli - ETHICS Why is it important to study the ethics of artificial intelligence? What is Ethics? What is the Ethics of AI? MORALITY Morality is the totality of opinions, decisions, and actions with which people, individually or collectively, express what they think is...

ETHICS, LAW AND AI Prof. Redaelli - ETHICS Why is it important to study the ethics of artificial intelligence? What is Ethics? What is the Ethics of AI? MORALITY Morality is the totality of opinions, decisions, and actions with which people, individually or collectively, express what they think is good or right. It is the totality of norms and values that actually exists in society. ETHICS Ethics is a systematic reflection on what is moral. It is not a manual with answers, a checklist of issues: it reflects on questions and arguments concerning the moral choices people can make. ARGUMENT The aim of an argument is first and foremost to justify or refute a statement. Argumentation is an activity that can be directed towards defending an opinion or attacking an opinion. Questions, orders and exclamations are not an argument. In general, an argument can formally be expressed as follows: A1, A2, A3 so B Where A1, A2 are premises and B is the conclusion. DESCRIPTIVE ETHICS AND DESCRIPTIVE JUDGEMENTS ​ Descriptive ethics is the branch of ethics that describes existing morality, including customs and habits, opinion about good and evil, responsible and irresponsible behavior, and acceptable and unacceptable action (how we actually act). It provides inputs for normative ethics. ​ Descriptive judgements describe what is actually the case (the present), what was the case (the past), and what will be the case (the future). It’s true or false. NORMATIVE ETHICS AND NORMATIVE JUDGEMENTS ​ Normative Ethics is the branch of ethics that judges morality and tries to formulate the normative recommendations about how to act or live. (how we should act). ​ Normative judgments are value judgments that indicate whether something is good or bad, desirable or undesirable; they often refer to how the world should be instead of how it is. (ex. “stealing is wrong”). “If the statement means that the law declares that taking bribes is illegal it is a descriptive judgment. If however, the statement means that bribery should be forbidden then it is a normative judgment”. The type of statement depends on the context or on what the speaker intends to say. VALUES Values are lasting convictions or matters that people feel should be strived for in general and not just for themselves to be able to lead a good life or to realize a just society.  ​ Intrinsic value: it is an objective value in and of itself. Something is good or important just because of what it is, not because it helps us get something else. We value it for its own sake. For example, happiness or kindness is valuable simply because it's good to be happy or kind. ​ Instrumental value: it is a means to realizing an intrinsic value. It is something we value because it helps us achieve something else that is good. It's like a tool or a means to reach an important goal. For example, money has value because it helps us buy things we need, but we don't value it just for the sake of having it. NORMS Norms are rules that prescribe what concrete actions are required, permitted or forbidden. These are rules and agreements about how people are supposed to treat each other. Values are often translated into rules so that it is clear in everyday life how we should act to achieve certain values. THREE KINDS OF ETHICS 1.​ DEONTOLOGICAL ETHICS (Kant) The type of ethics is characterized by the fact that it evaluates the ethical correctness of actions on the basis of characteristics that affect the action itself, rather than the outcome or consequences. Such a characteristic, for example, may be the intention with which an action is performed by the subject or/and the compatibility of the action with a particular formal principle. > ”Do not lie” or “do not kill” are seen as duties according to the universal social laws and are not based on what happens as a result or because they benefit you or others > One problem with this approach is that sometimes different duties can conflict with each other. For example, if you are asked a question where telling the truth could hurt someone, the duty to be honest conflicts with the duty not to harm others. 2.​ CONSEQUENTIALIST ETHICS Consequentialist theories determine the ethical correctness of an action or a norm solely on the basis of their consequences. For these approaches, what is important is only the result of the action. Therefore, an action in itself is not right or wrong, it’s only the consequence of action that is morally relevant. > A problem with this approach is how to determine the consequences of various courses of action in order to maximize some measure of utility. 3.​ VIRTUE ETHICS > Aristotele: Moral virtues are the desirable characteristics of people, whereas intellectual virtues focus on knowledge and skills. > A. Macintyre describes virtues as a certain type of human characteristic or quality that expresses a value that is worth striving for. (For example, being honest is a virtue because honesty is a value that helps build trust and respect.) > Classical View of Virtues: Moral virtues are character traits that make someone a good person or that allow people to lead good lives (acting on their basis was equally good for the person acting and for the person affected by their actions). AI ETHICS IS ABOUT HUMAN-AI INTERACTION AI ethics is about understanding and addressing the ethical challenges that come with artificial intelligence (AI) and how it affects our society. It’s not just about the technology itself, but also about how people use, interact with, and experience AI. This includes the decisions humans make when creating and applying AI, as well as the effects AI has on people's lives. The ethics of artificial intelligence is the branch of the ethics of technology specific to artificially intelligent systems. It is sometimes divided into a concern with the moral behavior of humans as they design, make, use and treat artificially intelligent systems, and a concern with the behavior of machines, in machine ethics. It also includes the issue of a possible singularity due to superintelligent AI. (Wikipedia) AI DEFINITION The European Commission's High-Level Expert Group on Artificial Intelligence defines AI as systems designed by humans (software and sometimes hardware) that are created to achieve complex goals. These AI systems can act in both the physical world (like robots) or the digital world (like software programs). AI works by 1.​ perception of the information from their surroundings by collecting and analyzing data 2.​ interpretation and reasoning on the knowledge, or process of it in order to make decisions 3.​ actions 4.​ learning and adaptation. AI systems can either use symbolic rules (following clear instructions) or learn by experience, creating a numeric model (this is often referred to as machine learning). As a scientific discipline, AI includes several approaches and techniques, such as machine learning (of which deep learning and reinforcement learning are specific examples), machine reasoning (which includes planning, scheduling, knowledge representation and reasoning, search, and optimization), and robotics (which includes control, perception, sensors and actuators, as well as the integration of all other techniques into cyber-physical systems). AI is both theoretical and pragmatic, it can be defined both as science and as technology. As a technology, AI has ethical and societal consequences. RATIONALITY AI achieves “rationality by perceiving the environment in which the system is immersed through some sensors, reasoning on what is perceived, deciding what the best action is, and then acting accordingly, through some actuators, thus possibly modifying the environment”. INTELLIGENT AGENT An agent is something or someone that acts. An agent is intelligent when: > Its action are appropriate for its goals; > It is flexible to changing environments and changing targets(!); > It learns from experience; > It makes appropriate choices given its perceptual and computational limitations. A DEFINITION OF TECHNICAL ARTIFACT Technical artifacts have three characteristics: 1.​ they have a practical function, 2.​ a physical composition, 3.​ instructions for use (that is a use plan) Technical artifacts have a peculiar nature: they are hybrid objects with physical and functional features. We can say that the physical features are intrinsic (or built into the object itself, such as shape, material and size), while functional features (what they are used for) are not, even if the two types of features are related. However, the functional features are also related to human intention or practices. > According to its nature, technical artifacts cannot be characterized fully in terms of their intrinsic physical properties alone. What distinguishes them from a mere physical object are some of its relational or extrinsic properties. USE PLAN A use plan is a plan that describes the proper use of a technical artifact, which will result (in the right context and with users with the right competencies) in the artifact fulfilling its proper function. >We do not classify natural objects as artifacts because they are not designed to fulfill a function, they do not have a (human) use plan. AI AS A SOCIOTECHNICAL SYSTEM ​ Kroes: Sociotechnical systems will here be understood as systems that depend on not only technical hardware but also human behavior and social institutions for their proper functioning ​ Traditional sociotechnical systems consist of three basic building blocks: 1) technological artifacts (tools or machines used) 2) human agents (people who operate, interact with or influence the system) 3) institutional rules (the laws, guidelines, or practices that people follow to use the technology effectively and safely). ​ The basic building blocks of an AI system: 1) Technical artifacts These are the AI software, hardware, or tools (like an AI-powered robot or a machine learning algorithm). 2) Artificial agents These are the AI programs themselves, which act like decision-makers or performers of tasks within the system (like a virtual assistant or an autonomous car). 3) Technical norms These are the rules or standards for how the AI system should function and be used, including programming guidelines, ethical standards, and safety regulation AI AS (NON-TRADITIONAL) SOCIOTECHNICAL SYSTEM Floridi and Sanders (2004) has explained such properties in the following terms: 1) Interactivity means that the agent and its environment (can) act upon each other. Typical examples include input or output of a value, or simultaneous engagement of an action by both agent and patient – for example gravitational force between bodies. 2) Adaptability means that the agent’s interactions (can) change the transition rules by which it changes state. This property ensures that an agent might be viewed as learning its own mode of operation in a way which depends critically on its experience. 3) Autonomy means that the agent is able to change state without direct response to interaction: can perform internal transitions to change its state. This property imbues an agent with a certain degree of complexity and independence from its environment. ​ Personal autonomy refers to the capacity to form personal values, goals, and ultimate ends. ​ Moral autonomy means the possibility of reflecting on one’s own moral principles or ethical convictions. ​ Rational autonomy seems to be prima facie achievable for artificial agents as well, as for Darwall, it is grounded solely on action on the basis of the weightiest reasons. ​ Agential autonomy: this form of autonomy consists of identifying a certain behavior as a “genuine action”, that is, an action not entirely determined by external factors. This may be represented by the artificial system’s ability to change internal states without external stimuli. HUMAN-TECHNOLOGIES INTERACTIONS Theories about technology: 1.​ INSTRUMENTALIST THEORY “The instrumentalist theory offers the most widely accepted view of technology. It is based on the common sense idea that technologies are ‘tools’ standing ready to serve the aims of users”. And because a tool or instrument “is ‘neutral,’ without valuative content of its own” a technological artifact is evaluated not in and of itself, but on the basis of the particular employments that have been decided by its human designer or user. Consequently, technology is only a means to an end; it is not and does not have an end in its own right. Technological artifacts do not have, have embedded in them, or contain values. Rather it is people who have values. “Guns don’t kill people, people kill people” ​ It is the use of technology, and not the technology itself, that is morally good or bad and thus has moral value. (Guns cannot use coercion or possession to make a person shoot): CRITICISMS Ihde argues that owning a gun transforms a person’s sense of self and agency, challenging the National Rifle Association's (NRA) view that firearm responsibility is solely a matter of human choice. Guns, designed for distant and life-altering action with minimal effort, inherently change how individuals perceive and interact with their surroundings. Gun possession amplifies boldness and shifts demeanor, leading to a world where people, animals, and objects are seen as potential targets. Ihde highlights how guns mediate human experiences by amplifying certain perceptions and reducing others, especially reducing perceived dangers and increasing responses to potential threats with violence. Latour adds that the relationship between a person and a gun transforms both the individual and the gun itself; the presence of the gun alters the person, and the gun takes on new significance when held. MEDIATION THEORY Mediation theory offers a more nuanced view by proposing that technologies mediate human-world relations, shaping how humans perceive and act in the world. According to this perspective, technologies are not simply tools; they influence how we experience reality. For example, when someone uses a gun, their sense of power, safety, and vulnerability changes, altering how they engage with others and their surroundings. Technologies are not neutral; they are not isolated from social and material contexts. As Melvin Kranzberg’s first law of technology famously states: “Technology is neither good nor bad; nor is it neutral.” Technologies have the potential to amplify certain aspects of human behavior and reduce others. In the case of a gun, it amplifies the user’s capacity for violence while reducing their perception of physical vulnerability. Don Ihde’s framework on human-technology interactions presents four distinct types of relations that illustrate how humans engage with technology and how technology mediates their experience of the world: 1.​ Embodiment Relations In embodiment relations, technology becomes an extension of the human body, allowing us to interact with the world through it without conscious awareness of the technology itself. The technology becomes “transparent” as we focus on the activity at hand, not the device enabling it. Example: When wearing glasses, we don’t focus on the glasses themselves, but rather on the world we see more clearly through them. Similarly, when driving a car, the car feels like an extension of our body as we navigate the road, feeling the distance between the car and the curb when parking. 2.​ Hermeneutic Relations In hermeneutic relations, technology is not transparent but becomes a tool through which we interpret or “read” the world. The technology is visible, and it helps us understand or decode aspects of the world by providing a specific interpretation. Example: A clock does not simply show the time; it shapes our interpretation of time by dividing it into measurable units. Similarly, using a thermometer to read temperature involves interpreting the information presented by the device, which mediates our understanding of the environment. 3.​ Alterity Relations In alterity relations, technology takes on the role of “the other,” becoming something we interact with directly. Here, technology is neither a tool to interpret the world nor a transparent extension of ourselves, but rather something that stands apart and engages with us almost as an independent entity. Example: Robots, especially humanoid or social robots, are experienced as quasi-others. They appear to us as more than just tools, behaving in ways that resemble human actions. We interact with robots in a way that feels similar to how we interact with other people or pets. Research shows that people empathize with robots and even hesitate to harm them, indicating that robots are perceived as having some form of agency or social presence. 4.​ Background Relations In background relations, technology shapes and mediates our experience, but it operates in the background, unnoticed unless it fails or needs attention. These technologies quietly influence our environment and actions without being the focal point of our attention. Example: A thermostat regulates the temperature of a room without us constantly thinking about it. It mediates our comfort, yet it remains in the background of our experience unless we need to adjust it or if it stops functioning properly. Cyborg Relations – Fusion Relations Cyborg relations represent a radical form of embodiment where the boundaries between human and technology blur, leading to a fusion of the two. In this relationship, technology not only enhances human capabilities but physically merges with the body, making it difficult to distinguish between the two. Example: Brain implants that directly interact with neural activity are an example of this fusion, where the technology becomes an inseparable part of the human being. This goes beyond using technology as a tool and represents a deep integration between human and machine. Immersion Relations In immersion relations, technology does not merge with the body but integrates with the surrounding environment. This environment becomes interactive, and human beings engage with it as an active participant. Example: A smart environment, such as a fully automated home with connected devices and sensors, is an example of immersion relations. The technology becomes part of the environment itself, and humans interact with it dynamically, leading to a seamless blend of physical space and digital augmentation. Multistability Multistability refers to the capacity of a technology to be used and understood in different ways across multiple contexts. A single piece of technology can be adapted to serve various purposes, depending on the situation or user. Example: An ultrasound machine used for obstetric purposes not only functions to visualize the fetus but also shapes the way parents and medical professionals understand and experience the unborn child. The technology influences decisions related to pregnancy, such as prenatal diagnosis and even considerations about abortion, by presenting the fetus as a potential patient. Verbeek’s Philosophy of Mediation It focuses on how technology actively shapes human experiences, decisions, and moral actions. It goes beyond the traditional view of technology as neutral tools and argues that technologies play an active role in shaping our understanding of the world and our moral decisions.  1.​ Technology as More Than Functional For Verbeek, technology does more than serve a functional role. Using the example of prenatal ultrasound, Verbeek explains that such technology doesn’t just make the unborn child visible; it also transforms the unborn child into a potential patient, turning the parents into decision-makers about the child’s future. This changes the nature of pregnancy, which becomes a process of decision-making. By making certain features of the fetus visible—such as predicting diseases—technology plays a role in how the parents experience their unborn child, shifting their understanding of it from an abstract concept to a medical subject. Technology thus redefines both the ontological status of the fetus and the moral decisions surrounding it. 2.​ Material Interpretation of Technology Verbeek argues that technology embodies a “material interpretation.” This means that technologies actively contribute to shaping human experience and prescribe specific behaviors. For example, in ultrasound imaging, the fetus is displayed as separate from the mother and enlarged to emphasize certain characteristics, leading to a moral and ontological shift in how the fetus is perceived. 3.​ Technological Intentionality Verbeek introduces the concept of technological intentionality, which refers to the way technology influences human actions. This intentionality is not reducible to the intentions of the designer or user; it emerges through the technology itself and affects human behavior in unforeseen ways. The effects of technology are not fully predictable or controllable, as they develop beyond the intentions of those who create or use them. For example, although the purpose of a smartphone is to aid communication, its influence on human behavior—such as addiction or changes in social interaction—is an emergent property of its use. 4.​ Designing Mediation Verbeek suggests that designers must anticipate the mediating roles that technologies will have in the future. Designing mediation involves understanding how a technology will influence human actions and decisions. However, this task is complex because the relationship between the designer’s intentions and the ultimate mediating effects of the technology is not always direct or predictable. There are three key actors in every interaction: ​ The human using the technology. ​ The artifact mediating the actions and decisions. ​ The designer who shapes the technology, whether explicitly or implicitly. Verbeek proposes the concept of moral imagination as a way to bridge the gap between the design context and the future use context, allowing designers to foresee how technology will mediate human experiences and moral decisions. Criticism: Peterson and Spahn’s Weak Neutrality Thesis Peterson and Spahn critique Verbeek’s philosophy, arguing that technological artifacts cannot be considered moral agents. While they acknowledge that technologies influence human actions and affect the moral evaluation of these actions, they maintain that artifacts are not morally responsible for these effects. This position is called the Weak Neutrality Thesis. It holds that technologies can shape how actions are evaluated but are not active participants in moral reasoning. They argue that technology is passive, and the active entity is the designer or user who decides how the technology is employed. For them, Verbeek’s view that technologies co-shape human existence blurs the line between human perception and reality. They believe that while technology may affect how we perceive the world, it does not actively shape reality itself. Example: Verbeek argues that different writing tools, such as fountain pens and word processors, actively shape how authors interact with their texts: these technologies guide or influence how they are used. He describes this influence as a kind of "intentionality," meaning that the technology promotes certain behaviors or practices. On the other hand, Spahn and Peterson acknowledge that technologies can influence how people behave, but they disagree with using the term "intentionality" to describe this influence. According to them, technological tools themselves (i) do not act as moral agents, (ii) cannot be held morally responsible for their effects, but (iii) can still impact how we morally judge the actions that involve their use. In other words, while technology can shape human behavior, it is not seen as having its own moral intentions or responsibility. Inherently political technologies The example of low-hanging overpasses on parkways in Long Island, designed by architect Robert Moses, illustrates how certain technologies or infrastructures can be used to enforce social inequalities. By constructing overpasses that were too low for buses, Moses limited public transportation access to the beach, effectively preventing African Americans, who often could not afford cars, from visiting these public spaces. This demonstrates how technologies can embody social and political relationships. Political theorist Langdon Winner explains that there are two main ways in which artifacts (technologies) can have political implications. 1.​ The first is when technologies are designed intentionally to support or promote specific social values or suppress others. 2.​ The second is when certain technologies are “inherently political,” meaning they naturally require or fit with certain political structures or relationships. Winner’s point is that when a society chooses a particular technology, it is not just choosing a tool but is also endorsing a certain political or social framework that can support their development and use. Examples include nuclear power, which necessitates centralized, state-level control and oversight, and the atomic bomb, which implies military power and significant political authority. Technological innovations as legislative acts Technological innovations, whether intentional or not, shape society in profound ways, similar to the impact of legislative acts. When a society decides on how a technology is structured, it significantly influences long-term behaviors, such as how people work, communicate, travel, or consume. These decisions are made within a context where different groups have varying levels of power and awareness. The most freedom to make impactful choices exists at the moment a new technology is first introduced. However, once these initial choices are embedded in physical infrastructure, economic investments, and societal habits, the flexibility to make changes becomes almost nonexistent. In this way, technologies, once adopted, become rigid and enduring parts of society, much like laws or political systems that set a foundation for public life. Therefore, the process of deciding on technologies should be given the same level of scrutiny as legislative or political decision-making. This includes careful consideration of infrastructure projects like highways, communication systems like television networks, and the small design details of new machines. The underlying social issues that bring people together or drive them apart are not only shaped by political institutions but also by the physical and technological structures embedded in society. Pitt against Winner Pitt argues against Winner’s perspective by challenging the idea that physical structures, like overpasses, can embody values on their own. He suggests that if we were looking at a schematic of an overpass and were asked to point out where "values" are present, we would be unable to do so. Pointing to the height measurement of the overpass, for example, only indicates a physical distance and not any inherent value. Pitt argues that claiming an object "embodies" values is only metaphorical. Even if Robert Moses had specific biases or intentions that led to the construction of low overpasses, it is those intentions and values that influenced the design, not the structure itself embodying them. In other words, while human intentions and decisions may shape technology, the technology itself does not possess or embody values independently (He stated that while there are many values involved in the creation of artifacts, it does not follow that the artifact is value-laden). ➔​ According to Pitt, Winner is pushing an ideology that advances the claim that a certain power structure or organization is responsible for technological results. Technological determinism Technological determinism is the belief that technology evolves independently, often without consideration for human needs, and has become so powerful that it dominates human life. This perspective suggests that technology has gained autonomy, leading to the idea that humans are now controlled by their own creations. This belief system leads to the philosophical stance that technology is the primary force shaping how we live, influencing our values, institutions, and social structures. There are two main types of technological determinism: 1.​ Hard determinism: This form of determinism attributes agency directly to technology itself or to its inherent characteristics. It posits that technology dictates the direction of society and is the driving force behind social change, making human influence negligible. 2.​ Soft determinism: This view acknowledges that while technology significantly influences society, it is not the only factor. It recognizes that technology’s development and impact are also shaped by social, economic, and cultural forces, indicating a reciprocal relationship between society and technology. Social Construction of Technology (SCOT) - Bijker and Pinch The Social Construction of Technology (SCOT), as proposed by scholars like Bijker and Pinch, emphasizes that the development and meaning of technologies are defined by social or interest groups. These groups influence what an artifact represents and how it is used, highlighting that technology results from the choices people make in response to the specific social, cultural, and economic conditions they experience. According to SCOT, technological development is not an autonomous or predetermined process but rather a socially constructed one. ➔​ However, an important limitation of this approach is that it focuses solely on how society influences technology, without considering how technologies, once established, may actively influence or impact society. In other words, while SCOT acknowledges the social processes that create technology, it does not account for the reciprocal effects that technologies can have on social structures, behaviors, and interactions. Actor-network theory - Latour Actor-network theory (ANT), developed by Bruno Latour, presents a view of reality as composed of networks of interconnected actors (or actants), which can be both human and non-human entities. These actors interact and form collectives that shape outcomes in social situations. A key feature of ANT is the principle of symmetry, which means that Latour’s analysis does not assume any inherent difference in status between human and non-human actors (equally significant in shaping outcomes). Latour also posits that artifacts serve as mediators within these networks. This means that they are not just passive objects but play an active role by being assigned programs of action, influencing how events unfold. >This distribution of action raises questions about the allocation of responsibility, as moral action is shared between human and non-human actors. Thus, ANT challenges traditional notions of responsibility by suggesting that it is distributed across an entire network of actors, blurring the lines between human intention and technological influence. Technological momentum - T.P Huges The concept of technological momentum, proposed by T. P. Hughes, blends elements of both technological determinism and social determinism to explain how technology and society interact over time. ​ Initially, when a new technology is introduced, society has significant control over how it is used and developed—this stage aligns with social determinism, where society shapes the technology according to its needs, values, and circumstances. ​ However, as the technology matures and becomes more embedded in society, it develops its own form of deterministic force. This stage represents technological determinism, where technology starts to influence and shape society in significant ways, becoming more autonomous and difficult to modify or steer. Hughes argues that over time, technology gains what he calls technological momentum. This means that as the technology integrates into the social and technological fabric, it builds inertia, especially within complex systems. The combination of its technological characteristics and its deep social integration makes it harder to change or redirect.ù AI systems can inherit and even amplify human biases when trained on historical data that reflect those biases. A clear example is in recruitment, where an AI assistant trained on past recruitment data quickly identifies patterns that were important for success, such as being white and male, because these were prevalent in the historical data. This leads the AI to prioritize those characteristics when selecting candidates, transferring and intensifying human biases into what is often mistakenly considered an unbiased, neutral system. Moreover, these biases become harder to identify and correct due to the opaque nature of AI decision-making. Biases in AI can emerge at various stages: ​ Design phase: During the choice of the training data. ​ Training data: If the data is unrepresentative or incomplete. ​ Algorithm itself: The model may inherently encode biases. ​ Post-training data: New data fed into the trained algorithm may be biased. ​ Spurious correlations: The AI may make decisions based on misleading patterns. ​ Developer influence: The biases of the developers who create the algorithm can be embedded in its design. Examples include: ​ using data that may be skewed towards certain demographics (e.g., based on white American males), resulting in biased predictions for a diverse population. ​ incomplete or of low quality datasets, further embedding inaccuracies. Developers' personal prejudices or assumptions can also affect outcomes, and biases may result from correlations that are not causally linked. For instance, if an algorithm learns that defendants with parents who have been imprisoned are more likely to be incarcerated, it might unjustly use that information to make harsher sentencing predictions, despite the absence of a causal link. ​ over-reliance on technology can lead to decisions being made without sufficient human oversight, perpetuating these biases further. Bias is a dimension of the decision-making, whereas discrimination describes the effects of a decision, in terms of adverse disproportionate impact resulting from algorithmic decision-making. The mirror view Should the data reflect reality or not? ​ Some argue that one should use a data set that mirrors the real world. The data may represent prejudices in society and the algorithm may model existing biases people have, but this is not a problem developers should be worried about. ​ Others argue that such a data set exists only because of centuries of bias, that the bias and discrimination are unjust and unfair, and that therefore one should change that data set or the algorithm in order to promote affirmative action. DOES AI HAVE POLITICS? The question of whether AI systems possess political characteristics revolves around whether these technologies inherently support certain political structures or ideologies. For instance, facial recognition systems have been widely implemented in China as a means of mass surveillance, serving to maintain order within an authoritarian framework. This raises the question of whether such technologies are inherently aligned with autocratic purposes. The early development of facial recognition by the US military (e.g., DARPA’s FERET program in 1993) hints that these systems may have been intended from the outset to serve control and enforcement functions. In some autocratic contexts, the use of such technologies for practices like racial profiling is not incidental but rather actively pursued, supporting the argument that certain AI systems may be more compatible with or designed for autocratic use. The concept of Value Sensitive Design (VSD) argues that technology can be deliberately embedded with moral values. VSD posits that through careful and intentional design, technological artifacts can reflect human values and promote ethical considerations. The goal is to ensure that technology development comprehensively incorporates human values at every stage, thereby making the resulting artifacts value-laden. In practice, this means that technologies can be designed to align with specific values, be they democratic, autocratic, or otherwise. For instance, if a facial recognition system is designed with transparency, accountability, and privacy in mind, it embodies those democratic values. On the other hand, systems designed without these considerations, or with the goal of maximizing control and surveillance, may naturally align with authoritarian uses. Ultimately, while AI systems may not have political intentions themselves, the values embedded in their design and use can reflect and support particular political structures. This highlights the importance of considering value-sensitive approaches when developing AI technologies to ensure they align with desired social and moral outcomes. Value Sensitive Design Value Sensitive Design aims at integrating three kinds of investigations: 1.​ Empirical investigations aim at understanding the contexts and experiences of the people affected by technological designs. This is relevant to appreciating precisely what values are at stake and how these values are affected by different designs. 2.​ Conceptual investigations aim at clarifying the values at stake, and at making tradeoffs between the various values. It strives to operationalize and measure the value of a design in one way or another in order to find a trade-off. 3.​ Technical investigation analyzes designs and their operational principles to assess how well they support particular values, and, conversely, to develop new innovative designs that meet particular morally values particularly well. Definition and types of values A value refers to a principle or standard used to assess the desirability or goodness of states of affairs, including technological artifacts. Values are associated with positive or negative evaluations of entities, while deontic notions like duties and norms relate to judging the rightness or wrongness of actions. For an entity to embody a value, it must provide reasons for a favorable response or behavior toward it. According to Van de Poel, if something is considered valuable, there are inherent reasons for positive attitudes or actions toward it, which originate from the object itself. 1.​ Intended values: These are the values that designers purposefully integrate into their design and aim to manifest in real-world applications. They represent the goals or principles that creators hope to see reflected in the technology’s use.  2.​ Realized values: These are the values that are actually expressed or brought into effect by the artifact when it is used in practice. Realized values show the outcome of the design in action within its practical context. 3.​ Embodied values: These refer to the potential of an artifact to express a value in the right circumstances. Embodied values are present because of the inherent properties designed into the object but may not necessarily be realized unless the artifact is used appropriately. Designated features and unintended features When discussing designated features and unintended features, it’s important to note that the designated features of a technical artifact are those that were purposefully included by the designers. These are the features that are meant to fulfill a specific function or achieve a particular outcome. However, artifacts can also have unintended features, which are side effects not intended by the designers. For example, while cars are designed for transportation, their pollution of the environment is an unintended feature resulting from their use, not a goal of the design. Incorporation of Values in Technical Artifacts: According to Van de Poel and Kroes, for a technical artifact x to embody a value V, it must meet the following criteria: 1.​ The designed properties of x must have the potential to achieve or contribute to V when used under appropriate conditions. This means x was created with the intention of fulfilling V. 2.​ The use of x must be conducive to achieving V. Additionally, these conditions must be connected in a meaningful way. Embodiment in AI In the context of AI, embodiment refers to the ability of artificial agents (AAs) to take on roles previously filled by humans within sociotechnical systems. Social institutions that govern human interactions can be transformed into computer code that manages the behavior and interactions of these artificial agents. Thus, while human behavior is regulated by social norms and institutions, the behavior of AAs is regulated by computer code, which acts as a translation of those social norms into technical norms. Van de Poel defines technical norms as the computer codes that regulate how AAs behave and interact. These norms can be created in two primary ways: 1.​ Offline design: Norms are predefined and programmed by human designers. 2.​ Autonomous learning: AAs are designed to autonomously discover, create, or spread norms through interactions with their environment or with other agents. Technical norms can also embody values. According to Van de Poel, a technical norm N embodies a value V if: 1.​ N is intentionally designed for V by human system designers. 2.​ The execution of N in the system supports or promotes V. Differences Between Human and Artificial Agents: Artificial agents (AAs) can embody values in their behavior and functioning, while it would be incorrect to say that humans embody values. Humans can hold, develop, or instill values in others but do not embody them in the way technical systems do. While humans can instill values into artifacts or AAs, artificial agents cannot instill values in other entities due to their lack of intentionality. However, AAs differ from traditional technical artifacts due to their autonomy, interactivity, and adaptability. This adaptability allows AAs to modify their behavior based on external stimuli and interactions, which can both enhance and challenge the realization of their embodied values: ​ Strengthening values: AAs can respond to new or unexpected situations, adapting to promote the values they were designed to embody. ​ Abandoning values: AAs can potentially adapt in a way that they no longer act in accordance with the values they originally embodied, effectively "disembodying" those values. An AA is said to disembody a value V if it changes itself such that it no longer promotes V under normal conditions, despite having been designed to embody V initially. The moral status of AI The moral status of AI involves addressing two main questions: 1.​ Moral Agency of AI: This concerns what an AI can or should be able to do from a moral standpoint. An agent is an entity capable of acting, with agency representing the expression of this capacity. Moral agency specifically refers to the ability to make moral decisions based on understanding right and wrong and being held accountable for those actions. 2.​ Treatment of AI: This pertains to how we, as humans, should behave towards AI. It relates to moral patiency, which is not about AI's ethics but our responsibilities and behavior toward AI systems. Moor’s Categories of Ethical Agents: James Moor proposed four types of ethical agents, each with different levels of moral capacity: 1.​ Ethical-impact agents: Machines that can be assessed based on their ethical consequences. For instance, a robot used in camel racing may have ethical implications due to its role and effects. The issue here is that any robot can be evaluated for its ethical impact. 2.​ Implicit ethical agents: Machines designed to avoid causing negative ethical effects. Their behavior is inherently ethical, as their internal design promotes ethical behavior or prevents unethical outcomes. This means they possess "virtues" to a limited extent. Ideally, all robots should be engineered to this standard. 3.​ Explicit ethical agents: Machines that use programmed ethical frameworks and reasoning processes, such as deontic logic, to make decisions involving duties and obligations. These agents actively reason about ethical issues as part of their function. 4.​ Full ethical agents: Machines that can make complex moral judgments and justify their decisions. Achieving this level of ethical capacity often implies the need for consciousness, intentionality, and free will. Full ethical agents would operate at a human-like level of moral decision-making and accountability. Moor identifies the goal of machine ethics as developing explicit ethical agents—machines capable of reasoning about and making ethical decisions based on structured ethical frameworks.  ➔​ We can give AIs principles, and machines might even be better than human beings at moral reasoning since they are more rational and do not get carried away by their emotions. The contra is that moral rules often conflict and morality cannot be reduced to following rules and is not entirely a matter of rationality, but the emotions may be indispensable for moral judgment. The fundamental QUESTION is whether consciousness is a necessary condition for possessing moral status. ​ Some authors argue against using consciousness as a moral criterion because we don't verify consciousness in humans before granting them moral consideration. This is linked to the "problem of other minds": we cannot directly access another being’s consciousness, making it fundamentally uncertain. Additionally, the concept of consciousness itself is vague and ambiguous, leading some to suggest it’s an unreliable basis for moral judgments. ​ Some others argue that machines cannot be moral agents because they lack essential qualities like mental states, emotions, and free will. For instance, Johnson contends that since machines are created and used by humans, only the humans possess the freedom needed for moral decision-making. ➔​ Johnson considers computer systems as sociotechnical systems: they have meaning and significance only in relation to human beings; through social practices, relationships, and knowledge. ​ However, others counter this by noting that while AI lacks its own intentions, it is intentionally designed and used by humans, giving it a kind of derived intentionality. This makes AI not morally neutral, as it can still have significant effects and be considered a part of the moral world due to its purpose and impact. Distinction between natural entities and human-made entities (Johnson) To support her thesis, Johnson distinguishes natural phenomena or natural entities and human-made entities. ​ For Johnson a natural object appearing in nature independent of human behavior is nature, a technology is a tool for a human ends. ​ While both natural objects and human-made objects have functionality, natural objects were not designed by humans: they do not have intentionality and could not be otherwise ➔​ Without this distinction, it would be impossible to recognize human impacts on the environment or to make informed choices about issues like climate change. The distinction allows us to understand human responsibility in shaping the world. Distinction between artifacts and technology (Johnson) ​ Artifacts, including computer systems, have been intentionally designed and poised to behave in the way they do (their functionality has been intentionally created). Johnson argues that artifacts gain meaning only within the context of human social activity. To identify something as an artifact, we mentally separate it from its social context. This act of separation abstracts the artifact from the real-world socio-technical system that gives it meaning and function. Removing them conceptually from this context overlooks the interconnected system of knowledge, practices, and human relationships they belong to.  ​ So technology is a socio-technical system, while an artifact is an abstraction from the system. Moral agent (Johnson) A moral agent is someone held accountable for their intended, voluntary actions. Responsibility does not apply to actions that were unintended or to consequences that were unforeseen. The key idea is that intentional behavior is driven by internal mental states (such as beliefs, desires, or intentions), which cause outward actions. This differs from involuntary behavior, which does not stem from conscious intentions. ➔​ All behavior has causes, only intentional actions can be explained by referencing the agent’s internal mental states, not just the external causes. The conditions for moral agency are: 1.​ Internal State: The agent has internal mental states (desires, beliefs, intentions) that provide a reason to act. 2.​ Outward Action: The agent performs a physical action or movement. 3.​ Causal Link: The internal mental state causes the outward action, aiming at a particular outcome. 4.​ External Effect: The action produces an effect in the external world. 5.​ Impact on a Patient: The effect must involve a recipient (the "patient"), who can be either helped or harmed by the action. Computer behavior meets several conditions of moral agency: 1.​ There is an outward action (condition 2). 2.​ This action is caused by an internal state (software or programming) (condition 3). 3.​ The action can produce an external effect (condition 4). 4.​ The effect can impact a moral patient (condition 5). However, computers lack freedom and true intentionality: they do not possess an intention to act, which is key to moral responsibility. This makes them unsuitable for moral appraisal, as the capacity for intentional action is central to moral agency. Machine freedom and computer intentionality Johnson explores the idea of machine freedom, noting that while neural networks may display a mix of deterministic and non-deterministic behavior, similar to humans, this does not equate to genuine freedom. The unpredictability of AI behavior stems from its learning processes, often beyond full programmer control. However, this kind of "freedom" is fundamentally different from human free will, which is tied to conscious intention and decision-making. The concept of computer intentionality helps clarify this distinction. Computer systems are designed with an inherent intentionality, meaning they respond in specific ways based on input. However, this intentionality is not self-generated but rather depends on the intentional acts of designers (who program the system's capabilities) and users (who provide the input that activates the system’s responses). > While machines can exhibit complex and seemingly autonomous behavior, their actions are rooted in and limited by the intentional frameworks established by human designers and users. Summary of Johnson analysis Johnson’s analysis of computer intentionality has two main points: 1.​ Dependence on Human Intentionality: Computer systems are fundamentally shaped by human intentions. Although they may operate independently of their designers and users in time and space, their behavior is rooted in the intentional actions of human creators. The system’s capabilities and responses are embedded by human intentionality. 2.​ Independent Behavior: Once deployed, computer systems can act without direct human intervention. Despite their reliance on human design, they exhibit a form of built-in intentionality that allows them to behave autonomously after being initiated. Johnson argues that computer systems, while not true moral agents (due to their lack of mental states and true intention), are still moral entities. Unlike natural objects, computers are closer to moral agents because their actions are purposefully crafted and used. When a computer system operates, three types of intentionality are involved: 1.​ Designer’s Intentionality: The creator embeds specific purposes and functionalities into the system. 2.​ System’s Intentionality: The computer exhibits behaviors based on its programmed responses and learned patterns. 3.​ User’s Intentionality: The user activates and guides the system through their input and choices. This triad highlights that the efficacy of a computer system arises from the interplay between the human designer’s intentions, the user’s actions, and the programmed responses of the artifact. Thus, computer systems are not neutral; they hold a significant role in the moral landscape because of their purpose, design, and impact. Johnson's argument concludes that computer systems cannot meet the core requirement of moral agency: they lack mental states and the capacity for intentions rooted in free will. Even if their states could resemble mental states, they do not have true intentions or autonomy. Therefore, they cannot be considered full-fledged moral agents. AI as Computation Artifacts or Sociotechnical Systems? Johnson distinguishes between two views of AI: 1.​ Narrow View: AI is seen purely as a collection of computation artifacts, without ethical considerations. 2.​ Broader View: AI is part of a sociotechnical system, involving both technology and the social context in which it operates. From this broader perspective, AI artifacts can have ethical implications, not inherently but as components of the wider system. AI Systems as Surrogate Agents Johnson and Powers propose that AI systems can be understood as surrogate agents, similar to human surrogate agents like lawyers or accountants. These systems act on behalf of humans, performing tasks assigned to them without having their own interests or values. ​ Surrogate Agents: Both human and computer surrogate agents operate with a third-person perspective, pursuing the interests of others rather than their own. They are obliged to act within the constraints of their roles and prioritize the interests of their clients or users.  ​ Key Difference: The main difference between human and computer surrogate agents lies not in their moral roles but in their psychology. While human surrogate agents have first-order interests and perspectives, computer systems lack these. They act purely on behalf of third parties, without personal intentions or desires. In essence, Johnson and Powers argue that while computer systems cannot be autonomous moral agents, they still function as important components in moral actions by serving as surrogate agents, fulfilling roles within a system of role morality. According to Johnson and Noorman (2014) there are three types of agency: 1.​ Causality: Artifacts possess causal efficacy, meaning they can actively influence events and alter states of affairs. However, this causal influence is insufficient to grant them moral agency. Despite their ability to effect change, artifacts only exert this influence in conjunction with human actions and decisions. In other words, their impact is mediated through human intent, design, and context, preventing us from ascribing independent moral agency to them. 2.​ Surrogate agent: Artifacts carry out delegated tasks on behalf of both their creators and users, as well as those who interact with them. In doing so, they extend the intentions of their deployers while also influencing the experiences of individuals who encounter them. 3.​ Autonomy: Humans are considered autonomous because they act based on reasons and deliberate intentions, placing their behavior beyond the realm of mere mechanical or material causality. In contrast, artifacts lack this form of autonomy, as they do not act for reasons but instead operate according to predefined rules or learned responses. >>Technologies don’t have autonomy and they can be moral agents just in the sense of surrogate agents and because they are part of socio-technical systems. Functional and Operational Morality - Wallach and Allen Wallach and Allen address a key challenge in AI ethics: How can we create an artificial moral agent? They explore the potential of machines equipped with both autonomy and sensitivity to values, which leads to the concept of functional morality (FM). This is distinct from operational morality (OM), and the difference lies in the machine’s capacity for ethical reasoning. ​ Operational Morality (OM) represents the most basic level of morality embedded in artifacts. Technologies with OM lack the ability to engage in ethical decision-making; instead, their ethical dimension is integrated through design choices that reflect certain values. For example, a gun with a childproof safety mechanism embodies OM. The gun does not assess moral implications but is designed with features that aim to prevent misuse, reflecting ethical concerns aligned with safety standards. ​ On the other hand, Functional Morality (FM) involves a higher level of moral capacity, enabling machines to assess and respond to ethical challenges. Unlike OM, FM requires systems to have the ability to evaluate the potential ethical ramifications of their actions. Examples of FM include autopilot systems and A I-based ethical  decision support tools, which are designed to make choices that consider ethical criteria, such as minimizing risk or ensuring passenger safety. Wallach and Allen argue that achieving FM is technologically possible if machines are equipped with autonomy (the ability to make independent decisions) and sensitivity to values (the capability to incorporate ethical principles in decision-making). Without these qualities, technology can only exhibit OM. At the lower end of the spectrum, we find systems like childproof guns, which lack both autonomy and sensitivity but still reflect ethical considerations in their design. At the higher end, systems with advanced autonomy and sensitivity can act as trustworthy moral agents, capable of evaluating the moral significance of their actions and behaving ethically in complex scenarios. According to Wallach and Allen, to produce the type of machine capable of decision-making, we can apply different ethical approaches and different ethical theories in the process of design. We can find three main different approaches: 1.​ top-down approach, 2.​ bottom-up approach, 3.​ Merging top-down and bottom-up approaches 1.​ Top-down approach The top-down approach to artificial morality involves defining a set of rules that can be transformed into algorithms, drawing from ethical theories like utilitarianism and deontology: ​ Utilitarianism: Morality is about maximizing overall utility, such as happiness or well-being. The best actions or rules are those that maximize aggregate utility. >Challenges with Utilitarianism ➔​ Computational Demands: Agents must evaluate numerous consequences of actions to determine the best option. ➔​ Evaluation Function: Difficulties in creating a computable function to balance present vs. future benefits or actual vs. potential outcomes. ➔​ Subjective Metrics: Assigning numerical values to subjective aspects like happiness or pleasure is problematic. ​ Deontology: Evaluates ethical correctness based on intrinsic features of actions, such as intentions or adherence to principles. Central to this approach are duties. >Challenges with Deontological Ethics ➔​ Internal Conflicts: Duties can clash (e.g., truth-telling vs. respecting privacy). ➔​ Rule Activation: Designers must ensure rules apply appropriately to situations and develop architectures for resolving rule conflicts. 2.​ Bottom-up and developmental approach The bottom-up and developmental approaches to artificial morality aim to create machines that mimic a child’s moral development, inspired by Alan Turing’s idea: "Instead of simulating the adult mind, simulate the child’s and subject it to education to develop into an adult brain." This approach focuses on developing discrete subsystems that cumulatively enable complex activities and autonomy. By experimenting with subsystem interactions, scientists aim to achieve higher-order cognitive faculties such as emotional intelligence, moral judgment, and potentially consciousness. The process is not explicitly guided by ethical theories, relying instead on emergent properties of the system. > Challenges with Bottom-Up and Developmental Approaches ➔​ Lack of Clear Goals: It’s challenging to define appropriate goals for evaluating choices and actions as contexts evolve. ➔​ Complexity with Multiple Goals: Systems directed at achieving one clear goal are simpler, but managing multiple goals or incomplete/confusing information is far more difficult. ➔​ Uncertain Objectives: Determining the appropriate ultimate goal for evolutionary artificial moral agents remains unclear. ➔​ Safety Concerns: The absence of preset safeguards makes these systems risky to deploy in real-world settings. 3.​ Hybrid model of top-down and bottom-up approaches Hybrid models have a base of rules or instructions, but then also are fed data to learn from as it goes, examples are Medethex and self-driving cars. Gunkel vs functionalist approahìch The functionalist approach has critical difficulties: 1.​ Testing: we will need some metric by which to evaluate whether or not a particular device is capable of making the appropriate moral decisions in a particular situation. 2.​ Anthropocentrism 3.​ Slave ethics Moral appearances of robots, Coeckelberg Coeckelbergh proposes developing robots that appear to be moral by mimicking emotions and behaviors in a way that humans interpret as signs of consciousness or morality. Humans do not require proof that another being has mental states or is conscious; instead, we rely on how the other appears and behaves. This interpretation allows us to interact with others as if they share our experiences and emotions, even if this cannot be confirmed. If robots are advanced enough to convincingly imitate subjectivity and consciousness, we would likely interact with them as if they were conscious beings. These robots would achieve "virtual" or "quasi-subjectivity," meaning their perceived moral standing would depend on how they appear and behave, rather than their actual ability to experience or understand morality. In this way, robots could become significant to us based on their appearances and the roles they play in social interactions, regardless of whether they truly possess consciousness. ➔​ Coeckelbergh argues that moral standing is not an objective quality inherent to a being but is shaped through human language and thought. It depends on human interpretation and subjectivity rather than existing independently. This perspective suggests that the way robots are perceived varies across different people, contexts, and cultures. The label "mere machine" is not necessarily the most accurate or meaningful way to define them, as their status is always open to interpretation. From this viewpoint, ethics shifts from focusing on what an entity inherently is (ontology) to how we perceive and understand it (epistemology). It emphasizes the subjective, relational aspects of morality, questioning how we construct and interact with entities like robots. Gunkel's Other-Directed Approach Gunkel critiques the "property approach," which assigns moral status based on inherent properties like intelligence or autonomy. Instead, he proposes an "other-directed" approach, inspired by Levinas’ philosophy, where ethics comes before questions about what a being is (ontology). In this view, moral status arises from interactions and our obligation to respond to the "other," rather than from the entity's intrinsic characteristics. This approach aligns with Coeckelberghs relational perspective, focusing on the significance of relationships and interactions in determining moral standing. Critical Flaws of Relationalism Despite its strengths, relationalism has significant limitations: 1.​ Lack of Clear Definition: Relational approaches describe social relationships but fail to provide a concrete definition of moral status. 2.​ Meta-Ethical Relativism: They risk extreme relativism, where moral status depends entirely on subjective perspectives, potentially undermining universal ethical principles. 3.​ The "Robinson Crusoe" Problem: Relationalism struggles to address cases where no social relationships exist, raising questions about the moral standing of isolated entities. Mindless Morality (Floridi and Sanders, 2004) Floridi and Sanders propose that moral agents do not necessarily need to possess consciousness or emotions to be considered moral. Instead, moral agency can be based on functional criteria. A moral agent must meet the following conditions: 1.​ Interactivity: The ability to respond to external stimuli by changing its state. 2.​ Autonomy: The capacity to change its state without direct external input, demonstrating independent action. 3.​ Adaptability: The ability to modify its own rules or behavior based on experience or changing circumstances. An agent's morality is evaluated based on its actions: ​ Morally good: When all actions meet or respect a certain moral threshold. ​ Morally evil: When at least one action violates that threshold. This framework defines moral agency without relying on subjective qualities like consciousness, focusing instead on measurable actions and capabilities. Sullins’ Three Requirements for Full Moral Agency Sullins identifies three key conditions for a robot to be considered a full moral agent, capable of moral responsibility: 1.​ Autonomy: A robot must operate independently from programmers, operators, or users, without direct external control, in an engineering sense. If the robot can  perform tasks effectively and achieve its goals on its own, it demonstrates "effective autonomy." When this independent action results in morally significant outcomes (good or harm), the robot can be said to have moral agency. 2.​ Intentionality: While proving a robot’s "true" intentions is unnecessary (and impossible, even for humans), its behavior should appear intentional. If the robot's actions—shaped by programming and its environment—are sufficiently complex and result in deliberate-looking outcomes (moral good or harm), it qualifies as a moral agent. The focus is not on whether the robot genuinely has free will but on whether its behavior mirrors intentionality at a level comparable to human interactions. 3.​ Responsibility : A robot can be seen as morally responsible when its actions fulfill roles that carry specific duties or responsibilities. For instance, if a robotic caregiver behaves in a way that aligns with its assumed duty to care for patients, and this is the only way to make sense of its actions, it can be ascribed moral responsibility. This recognition depends on the robot's capacity to consistently fulfill its role in a socially meaningful way. Example: Robotic Caregivers Sullins provides the example of robotic caregivers designed to assist the elderly. A human nurse is clearly considered a moral agent due to their autonomy, intentional actions, and sense of responsibility. Similarly, a robotic caregiver can be viewed as a moral agent if it: 1.​ Operates autonomously, independent of direct human control. 2.​ Displays intentional behavior through complex programming. 3.​ Understands its role and responsibilities in the healthcare system, caring for patients as part of its duties >> Robots can be considered moral agents when they exhibit autonomy, intentionality, and responsibility to a sufficient degree. At a reasonable level of abstraction, a machine that appears to act with autonomous intentions and fulfills social roles responsibly qualifies as a robust moral agent. Sullins argues that as robotic technology advances, future robots with highly complex interactivity may even approach or exceed the moral status of humans. However, even current, simpler robots can be seen as moral agents in specific contexts and deserve moral consideration based on their roles and behavior. Postphenomenological reinterpretation of moral agency ➔​ Intentionality: indicates the directedness of human beings towards reality. Human intentionality is mediated by technological artifacts. ➔​ Freedom (nondeterministic nature) is the capacity of technology to go beyond the use plan, the human intentionality. ➔​ Composite Intentionality: “there is a central role for the ‘intentionalities’ or directedness of technological artifacts themselves, as they interact with the intentionalities of the human beings using these artifacts” (Verbeek) AI is a moral mediator with composite intentionality. Andreas Matthias: a responsibility gap Matthias argues that modern machines with machine learning capabilities can autonomously modify their operational rules and act without human intervention. This creates a problem of responsibility attribution: traditionally, either the operator or manufacturer was held accountable, or no one in cases of no identifiable fault. However, these traditional approaches are no longer adequate for certain autonomous machine actions. Such situations conflict with our sense of justice and society's moral framework, as no one has enough control over the machine's actions to take responsibility for them. 1.​ Active responsibility: is responsibility before something has happened. Mark Bovens mentions the following features of active responsibility: 1.​ Adequate perception of threatened violations of norms; 2.​ Consideration of the consequences; 3.​ Autonomy, that is, the ability to make one’s own independent moral decisions; 4.​ Displaying conducts that is based on verifiable and consistent code 5.​ Taking role obligations seriously. 2.​ Passive responsibility is after something undesirable has happened. Four conditions are required to assign passive responsibility (accountability): 1.​ wrongdoing 2.​ causal contribution 3.​ foreseeability 4.​ freedom of action Under what conditions can one take responsibility? The use of autonomous robots challenges the traditional conditions for attributing responsibility as defined by Aristotle: 1.​ Control Condition: Autonomous robots modify their behavior independently, limiting human control and making their actions not fully traceable to human decisions. 2.​ Epistemic Condition: Autonomous robots act unpredictably, making it hard for humans to fully understand or anticipate their actions and consequences. However, demanding sufficient human control may fail in fast-reacting systems like autonomous weapons or high-frequency trading, or in systems that override human decisions, like future autopilots. >>While banning such systems is suggested, global competition may make this impractical. In many technologies, typically many people are involved in the development and use of autonomous robots, which makes it difficult to hold one individual responsible. >> Verbeek argues that even if technologies influence human actions, it does not mean that technologies themselves should be held morally accountable. Unlike humans, technologies lack the capacity for moral agency—they cannot deliberate, understand ethical principles, or take responsibility for actions. Autonomous robotics complicates responsibility further due to the involvement of "many hands" (numerous contributors) and "many things" (varied hardware and software components). These interconnected systems create uncertainty in identifying failures or assigning responsibility. For example, in self-driving car accidents, it is often unclear which component failed, how it caused the issue, or who should be held accountable. Epistemic Conditions of Responsibility Responsibility requires awareness; ignorance undermines it. Aristotle identified several forms of ignorance, including: ​ Not knowing what one is doing or the tools one is using. ​ Lack of awareness of the moral significance, consequences, or alternatives of actions. Applied to robotics, this means developers and users must understand the technology they are creating or employing and its ethical implications. Ignorance about tools, especially advanced systems, results in moral failures. Bridging the "knowledge gap" is essential to ensure accountability and avoid these failures.. The Responsibility and Knowledge Gap A "responsibility gap" emerges due to insufficient understanding among developers, users, and stakeholders of autonomous systems. Key issues include: ​ Developers’ incomplete knowledge: Many contributors may not fully grasp the system’s behavior or interactions. ​ Users’ lack of understanding: End-users often do not comprehend the underlying models, assumptions, or decision-making mechanisms. This gap is morally problematic because neither party fully understands their actions or their consequences. To address this, developers and users must actively improve their knowledge of the system’s operations, potential impacts, and moral responsibilities. Challenges of Predictability and Opacity Autonomous systems introduce additional difficulties: ​ Unpredictability: Robots may act in unforeseen ways, leading to unintended consequences. ​ Opacity: Black-box algorithms, such as neural networks, obscure the reasoning behind decisions. Relational Approach to Responsibility Responsibility extends beyond control and knowledge—it requires answerability. Developers and users must: ​ Be willing and able to explain and justify their actions and the system’s behavior. ​ Remain accountable to those affected by the robot’s actions, directly or indirectly. For instance, companies deploying self-driving cars should be answerable to victims and families of accidents. Developers and users must ensure the technology is designed and utilized to maintain accountability to all beneficiaries or those harmed by its outcomes. Key Insights: ​ Responsibility in autonomous robotics is hindered by a "knowledge gap" and challenges in assigning accountability. ​ Developers and users must prioritize awareness and transparency to mitigate moral failures. ​ Answerability is essential for maintaining ethical accountability in the development and use of these technologies.  ​ Collective Responsibility: Responsibility for autonomous robots is often distributed across a network of agents (both human and non-human). This concept is split into: ○​ Collective Responsibility: A unified moral agent (e.g., a corporation or state) bears responsibility. ○​ Shared Responsibility: Individual members of a group share responsibility for harm, whether through direct actions or group affiliation. 2. Challenges to Instrumentalist Responsibility ​ Instrumentalist View: Treats robots as mere tools. ​ Critique by Gunkel: Innovations like self-driving cars, AlphaGo (an AI for playing Go), and social robots challenge this view. For instance: ○​ Autonomous systems like self-driving cars replace the human driver, not just the car as a tool. ○​ AI victories, like AlphaGo defeating a human, raise questions about who receives recognition for achievements (human or AI?). 3. Moral Patiency ​ Moral Treatment of AI: ○​ Debate: Do we owe moral obligations to AI? Joanna Bryson argues no, as robots are tools and property. ○​ Counterpoint: People often empathize with robots and feel discomfort in mistreating them (e.g., destroying or harming robots like Boston Dynamics’ Spot or HitchBot). ○​ Kant’s Argument: Mistreating AI is not wrong because of harm to the AI but because it reflects poorly on human moral character. 4. Ethical Dilemmas with Humanoid Robots ​ Uncanny Valley: Humanoid robots that imperfectly resemble humans can evoke unease or revulsion. This raises ethical questions about: 1.​ Whether humanoid robots (like androids) should be developed at all. 2.​ Restricting their use for vulnerable groups (e.g., children, elderly). 3.​ Transparency in design: Robots should not deceive users into thinking they are human. Designers should clearly communicate a robot’s machine-like nature. 5. Practical and Ethical Cases ​ Mistreatment of Robots: Cases like kicking Spot (Boston Dynamics) or destroying HitchBot highlight public discomfort with harming robots, even when they are clearly machines. ​ Design Ethics: Proposals for "honest design ethics" suggest designers should acknowledge robots as machines and frame interactions as staged performances, akin to stage magic. 6. Indirect Moral Standing (Kantian Perspective) ​ Kantian View on Animals: Harm to animals damages human moral character, rather than directly violating the animal’s rights. Similarly: ○​ Mistreating AI is wrong, not because AI has rights, but because it erodes human kindness and moral integrity. Key Themes ​ Accountability: Responsibility in developing and using AI should ensure accountability for actions and consequences. ​ Ethics of Interaction: Human behavior toward robots reflects on broader ethical values, even if robots lack rights. ​ Transparency: Honest and non-deceptive design is crucial for ethical robotics development. ​ Moral Implications: Debates about moral obligations toward AI parallel long-standing philosophical questions about rights and responsibilities toward animals and other entities.