PP240H - Ethics and AI PDF

January 6th, 2025 PP240: Ethics and AI 1. What is AI? → Technology that allows computers to simulate human comprehension and problem- solving → Answer generated by Google AI → When you simulate something, A is NOT B, it is an imitation The Goa...

January 6th, 2025 PP240: Ethics and AI 1. What is AI? → Technology that allows computers to simulate human comprehension and problem- solving → Answer generated by Google AI → When you simulate something, A is NOT B, it is an imitation The Goal: Creation of Artificial Agents → This is what makes AI unique in the whole history of technology → We are all agents we act for reasons/ for purposes, given the environments we are in and the inputs we receive from those environments → Extends human capabilities = makes our lives more efficient An Early Warning → “If we use, to achieve our purposes, a mechanical agency with whose operations we cannot interfere effectively… we had better be quite sure that the purpose put into the machine is the purpose which we really desire.” (Norbert Wiener, “Some Moral and Technical Consequences of Automation” 1960) Questions From Adjacent Disciplines → 1. Philosophy: → Can formal rules be used to obtain valid conclusions (logic) → How does the mind arise from a physical brain? → What is knowledge and where does it come from? → What is the connection between knowledge and action? → What ethical issues are raised by the ways in which technology mediates our relation to the rest of the world? → 2. Mathematics: → What are the formal rules for drawing valid conclusions? → What can (and cannot) be computed? → How do we handle uncertainty rationally? → 3. Economics:  → How do we make decisions that reflect our preferences? → How should we do this when others do not agree with us? → How do we assess payoffs in the far future? → 4. Neuroscience: → How do brains engage in information-processing? → 5. Psychology: → How do humans think and act? → 6. Computer Engineering: → How can we build efficient and powerful computers? → 7. Control Theory and Cybernetics: → How can machines operate under their own control? → How much autonomy can we give machines? Can they truly have autonomy? → 8. Linguistics: → What is the relation between language and action? → This is a project that is so complicated that it necessarily draws in all of these complex disciplines to understand the project at hand Key Developments: Machine Learning, Deep Learning, Neural Networks, Gen AI → 1980s: Artificial Intelligence – Simulate with a computer something that will match or exceed human intelligence (infer, reason, etc.) → 2010s: Machine Learning – The machine is learning, it is not necessarily programmed, it predicts based on patterns → 2010s: Deep Learning – Neural networks (many layers, complicated) – simulate ways the human brain operates, → Generative AI – Foundation models – large language model, where we make predications where we see certain words, we can predict what other words will come next (ex: auto complete: predicts the next work while you’re typing) → It is generative because it is generating new ideas (i.e., music notes) → Chat bots – → Deep fakes – The AI Mirror → Gen AI reflects back to us only what we have deposited in data 2. Key Applications → Robotic vehicles → Legged locomotion → Autonomous planning and scheduling → Machine translation → Speech recognition → Recommendations → Game playing → Image understanding → Medicine → Climate science → Identifying terrorist threats in data → Assigning risk scores to defendants in bail, sentencing and parole evaluations → Assessing the likelihood that you fit into a company’s corporate culture → Assessing how close a stranger is to your romantic preferences → Assessing your kid’s chances of thriving at a private school → Automating lethal weapon systems 3. Ethical Issues → Technological unemployment → Limits of the ‘quantified self’ → Biases in algorithms → Opacity in machine decision making, leading to problems of justification → How morality can be programmed → Will AI itself become sentient? → Will we have duties to these machines? → Should they have rights? 4. Existential Issues → Impact on conceptions of sex and love → Impact on human self-understanding in a world that contains beings more intelligent than us → Requires a complete reconception of our place in the world (intelligence hierarchies) → Impact on human agency in a world where more and more of our functions are taken over → Grasping that AI is essentially a mirror of us and can create nothing new → How we understand meaning in work, love, war, etc.  → Facing the question: what if we ourselves are simulations? January 8th, 2025 1. 4 Moral Theories → Ethics is responsible for answering justificatory questions in a principled way → If machines enact our purposes, then the issue of ethics comes up A Few Preliminary Points → Ethical theories provide general principles dividing actions up into the permissible (morally speaking) and the impermissible (morally speaking) → There are 2 functions of ethics: to prevent harms and to promote human flourishing (fulfill our nature = to be moral creatures) → Ethical considerations are constraints on what people otherwise would do based on their desires 1. Utilitarianism: Jeremy Bentham (1748-1832) → Definition: A permissible action is one that will lead to the greatest balance of pleasure over pain among all sentient beings whose interests will be affected by a proposed course of action → What matters morally is if something can suffer, i.e., if it can feel pleasure or pain → If it can suffer then it has moral standing and its interests matter → Pleasures and pains can be quantified → Key slogan: “The greatest happiness for the greatest number of affected beings” → This is a consequentialist theory = What counts morally are the consequences of my actions, not my intentions, principles, adherence to religious practices, or desires, which are completely irrelevant → Animals are on equal moral standing with human beings 2. Deontology: Immanuel Kant (1724-1804) → Acting out of a sense of moral purity, acting for the right reasons based on the situation The Categorical Imperative (I) → Act only according to those rules that anybody in similar circumstances could act by → Don’t make an exception of yourself Example of Immorality: Lying → Imagine someone who borrows money with no intention of paying it back → According to Kant, lying is never permissible The Categorical Imperative (II) → Treat other people as ‘ends in themselves’ rather than as mere means to the fulfillment of your ends or desires → Don’t ‘use’ people → Kant invented the term ‘autonomy’ in the 1800s → Auto = self, Nomos = law → Autonomy = Give the law to yourself → Informed consent is respecting one’s interest and their autonomy → The consequences of that action don’t matter since your hands are ‘clean,’ you’ve done all you can as a moral agent → Philosophical foundation of the doctrine of human rights The Key to Both Forms → Intentions, not consequences are what matter morally 3. Virtue Ethics: Aristotle (384BCE – 322BCE) → Aristotle on virtue: The life of reason is the life of virtue → What is a virtue? A state of character that causes us to act in specific ways on specific occasions → Examples of virtues: courage, liberality, modesty, kindness, pride, justice, etc. → So, for example, the courageous person will do what courage demands in situations that call for it (in battle, say) → Courage demands different things from different individuals (Burning building example) Doctrine of the Mean → Definition: Every virtue is a middle point between 2 morally impermissible states (the extremes) Subject Defect Mean Excess Managing fear Cowardice Courage Rashness Managing honour Undue humility Proper pride Vanity 4. Social Contract Theory: Thomas Hobbes (1588-16679) → Amid the English Civil War, a state of nature became severed → A war of all against all Hobbes’ Claims → Civil society emerges out of the state of nature → State of nature is a ‘war of all against all’ → Life is ‘nasty, poor, brutish and short.’ → To escape this state, people ‘sign a contract’ (metaphorically speaking) to hand over their right to attack others to an all-powerful authority (The Leviathan) → Why do we sign? To avoid being victimized by our neighbours → Key: The authority invents right and wrong, just and unjust The Heart of the Theory → 1.) An action is permissible if it is the product of an agreement among rational agents → 1.) If it is not, it is impermissible → In Hobbes’ theory, a rational agent is anyone who is tired of being victimized → 2.) Morality is, therefore, a fundamentally social phenomenon Summary of the 4 Theories → An action is permissible only if it… → 1.) …produces the best consequences (utilitarianism) → 2.) …is motivated by the right reasons (deontology) → 3.) …is motivated by the proper character traits (virtue ethics) → 4.) …is sanctioned by an agreement among rational agents (social contract theory)  January 13th, 2025 Title 1. Social Perspective on Learning → What are we trying to do? Build mavhines that are “sensitive to ethical concerns” (45) → What does this mean? → A capacity to respond to robustly to ethically salient features of actions, situations, outcomes, and other agents → Railton think this capacity is at the heaet of what it means to be a social agent → It follows that we are trying to build ideal social agents An Analogy → We already recognize many non-human agents: governments, corporations, institutions, political parties, unions, etc. → Not just human beings despite them being composed of human beings → We share social space with these entities → We – individuals → We are part of those social agents and that’s why we share social spaces → Negotiate our way to the world with respect to the presence of other entities → They have no internal states (there is nothing it is like to be one of them) → They do not have internal states → Something like to be one of us – to be a bat (Nagel example) → They can have goals and ‘preferences’ that are not reducible to those of their individual members → The goals of university is not reducible to the president of the university, the members, deans, or profs → They behave both egocentrically and non-egocentrically → Not everything they do need be explained in terms of egocentric preferences and goals Social Contract Theory → RECALL: a course of action is justified to the extent that it emerges from an agreement among rationally informed stakeholders → Notice that this presupposes that each of us is responsible to ethically relevant features of our situation  → ‘Situations’ always includes other agents → Not just about objects in the world but other people with interests A Question and a Distinction → How should we understand this capacity? → Interested in agreeing with others and sensitive to ethical salient features of the situations → Two options: → It is an add-on module to general intelligence → So: we can be intelligent and either have this module or not → It is a deep feature of general intelligence → So: being intelligent tends to come with this capacity The Appeal of the Deep Feature View → Railton considers empirical research on 2 anti-social types lacking this capacity to some extent → Psychopaths: Show “serious deficits in attention, impulse control, and ability to accurately represent likely negative future outcomes” (48) → Machiavellians: Tend to adopt an egocentric “economic approach” to interpersonal relations which erodes trust → Conclusion: Perhaps ideal cooperative behaviour is a deep feature of general intelligence → Where did it come from? 2. Developmentalism in Ethics → Concept of “priors” → 2 features of early learning → Product of our biological heritage, innate structures which support specific capacities → 1. Curiosity: infants display a “form of internal motivation to learn above and beyond any more specific purpose they might have” (53) → 2. Trust: infants learn a default reliance on their own faculties: perception association, memory, etc. This becomes honed over time (it becomes less ‘blind’) → Key: priors are purely epistemic, but Railton argues that they are the foundation of ethical development 2 Important Accomplishments → An intuitive physics: figuring out how cause and effect work in the world of objects → An intuitive psychology: figuring out how other agents work, which involves developing a ‘theory of mind’ → Theory of mind – to be able to say that there are other objects in the world with internal states (a mind) → Key: The infant is learning how to model and predict the behaviour of things in its world. This allows the child to become increasingly autonomous in the social world The Deep Feature View → Infant learning builds an epistemic map of the world that simultaneously builds an intuitive ethics → Two capacities are key to this process: → 1. An attunement to, and ability to predict, the interests of other members of the social group → 2. An ability to evaluate novel social situations without explicit instruction → Doesn’t guarantee that even if I develop all of these things I will care about others (Be cautious with Railton) → Railton’s paper does not account for this possible gap What About Adults? → Railton’s Argument: Sound ethical judgement is largely a matter of intuition, not of rule- application → How to prove this? A bit of trolley-ology Trolley Problem: Switch → Intervention, pull the lever to kill 1 instead of 5 Trolley Problem: Footbridge → Push a person off the bridge to save 5 Railton’s Findings  → A vast majority think the switch should be pulled in the “Switch” → A vast majority think the person should not be pushed in the “Footbridge” → Ethical theory cannot account for this asymmetry: → Utilitarianism: Kill one person in both cases (Greatest good for the greatest amount of people) → Deontology: Kill nobody in either case (Violates categorical imperative to kill) → Both moral frameworks fail to capture what people think about the paradigm of these ethical issues Railton’s Hypothesis → The asymmetry can be explained as an expression of our intuitions about what sorts of people we want to be and to have around us → Trust is foundational here: → A majority indicate they would generally trust a person who reported pulling the switch in “Switch” → A majority indicate they would generally not trust a person who reported pushing the person in “Footbridge” Test This on Yourself → How would you feel about reporting your action to the dead individual’s family after the fact? → Railton’s suggestion: → In “Switch,” maybe you’d feel guilty but expect the family to understand → In “Footbridge,” maybe you’d feel shame and not expect the family to understand (unless you were a psychopath?) About Ethical Intuitions → They are developed alongside our cognitive-epistemic capacities, beginning in infancy (this is the deep-feature view) → The key to this is developing our predictive capacity: we need to be able to reliably predict the future based on relatively scant data → All of this evolves in an intrinsically social context → Because social life is messy, ethical intuition is too  → It is difficult to codify (arrange according to a plan or system) Implications for AI Ethics → Historically, machine ethics involved determining how to code fully formulated rules of conduct into algorithms Railton’s Challenge → But what if instead, we thought of AI’s akin to infants, learning how to behave ethically from the ground up? → This fits deep machine learning better → AI’s are trained on large data sets with a view to making them increasingly autonomous → Data are not rules! Rather, the AI is learning to make inferences and predictions on the basis of raw information, just like the infant An Alternative Paradigm for AI → We should think of ourselves as teachers rather than technicians (Large Language Models are little babies!) → A teacher is a focal point for communal norms but extends past the teacher → We should understand that we are trying to create reliable social partners rather than mere tools → As such, we should develop AI’s “priors”: their curiosity and, especially, capacity for trust and trustworthiness → We should recognize the dangers of this enterprise: that we might create psychopaths or Machiavellians (focused overwhelmingly on the pursuit of their own interests, rather, we want them to always be subservient to our needs/desires) →