Philosophy and Ethics for AI Week 1 Lecture 1

Ethics

The study of principles and values used to judge actions as right or wrong

Philosophy

The study of fundamental questions, such as: What is knowledge, and how do we acquire it? What is a mind (and could a machine have one)?

Philosophy of X

Asks about X’s concepts, methods, and assumptions, but are not answerable from within X alone

Are Philosophical Questions Scientific Questions?

Somewhat related to science, but too fundamental to belong to science. Philosiphical questions do influence how we think about science and helped with the progress of science.

Two fields of AI

Engineering: building systems that perform tasks we associate with intelligence. Science: understanding the principles that make such performance possible.

Four classic ways to define AI

Thinking humanly: modeling human cognition (cognitive science, psychology). Acting humanly: behaving like a human (e.g., conversation, perception, manipulation). Thinking rationally: correct reasoning (logic, proofs, formal inference). Acting rationally: choosing actions that achieve goals well under uncertainty (decision theory).

6 forms of intelligence

▶ Perception: extracting structure from sensory input (vision, speech). ▶ Learning: improving with experience, generalization beyond training data. ▶ Reasoning: drawing conclusions, planning, causal inference, explanation. ▶ Language: understanding and generating meaningful communication. ▶ Action: controlling systems (robots), interacting with people, tool use. ▶ Social intelligence: modeling others, norms, cooperation, deception.

Weak and Strong AI

▶ Weak AI: systems that simulate intelligent behavior in particular domains, without assuming genuine understanding. ▶ Strong AI: systems whose intelligent behavior is taken to reflect real understanding, minds, or consciousness.

AGI

Artificial General Intelligence: broad competence across many domains, with flexible transfer, rapid learning, and the ability to solve new problems without task-specific redesign.

Difference strong AI and AGI

▶ AGI is mainly a claim about breadth and flexibility of competence. ▶ Strong AI is mainly a claim about having a mind (understanding, consciousness).

How to answer the question if machines can think?

▶ A behavioral perspective argues that if it behaves as if it understands, treat it as understanding. ▶ An internalist perspective argues that behavior is not enough: we need the right internal structure (or the right causal connection to the world).

Difficulties with tests for consciousness in AI?

You can only see output, not what happens inside. Maybe understanding emerges from performance.

What is the aim of AI?

Technological aim: build systems that exhibit intelligent behavior (useful, reliable, safe). ▶ Scientific aim: understand intelligence (representation, learning, reasoning, agency).

How is AI implemented?

▶ Virtual: software agents (search engines, recommenders, chatbots, decision support). ▶ Physical: embodied systems (robots, autonomous vehicles, drones).

Approaches to AI

Symbolic AI (“good old-fashioned AI”) and Subsymbolic AI (machine learning / statistical AI)

Epistemology

Subdomain of philosophy about knowledge

Epistemology connections to symbolic AI (3)

▶ Knowledge Representation and Reasoning: How can we model knowledge explicitly in a way machines can process? ▶ Epistemic Logic: Understanding how AI systems can reason about what they (and others) know. ▶ AI verification: Ensuring systems behave consistently with encoded knowledge.

Epistemology connections to subsymbolic AI (3)

▶ When a model generalizes well, what exactly has it learned: structure or surface patterns? ▶ Three recurring epistemic challenges: ▶ Distribution shift: training conditions differ from deployment conditions. ▶ Adversarial fragility: small input changes can cause large output changes. ▶ Explainability: what kind of explanation is needed for justified trust and responsibility?

Philosophy of Science examples

▶ What makes a scientific explanation valid? ▶ How do we verify or falsify theories? ▶ What is the role of causality in science?

Philosophy of Science connection to AI

▶ What constitutes a good explanation for AI predictions? ▶ How do we design algorithms to explore hypotheses (explore vs exploit in RL)? Inductive biases in models: What assumptions guide learning? ▶ How can we go beyond correlations to uncover true causal relationships?

Philosophy of Mind connection to AI

▶ Weak vs strong AI debates: tool vs mind. ▶ Embodied AI: the role of action and environment. ▶ Neuroscience-inspired models: what can (and cannot) be learned from brains? ▶ Language models: what is the relationship between linguistic competence and understanding?

Philosophy of Mind examples

▶ What is a mind, and how is it related to the brain/body? ▶ Is consciousness necessary for intelligence? ▶ Are mental states computations, or something else? ▶ Is language necessary for thought?

Metaphysics examples

▶ What is the fundamental nature of reality? ▶ What exists (numbers, time, possibilities)? ▶ Is time an illusion? What about causality?

Metaphysics connection to AI

▶ Simulation Hypothesis: Are we living in a simulation, and can AI help detect it? ▶ Representation Learning: Can AI find a universal representation of the world, or is all representation anthropocentric? ▶ Ontology in AI: How do we encode ‘what exists’ in a machine-readable way?

Ethics examples

▶ What is the nature of morality? ▶ How can we distinguish good actions from bad ones? ▶ What is responsibility? ▶ Are ethical principles universal or relative?

Ethics connection to AI

▶ Value Alignment Problem: ensuring AI systems align with human ethical principles (e.g., trolley problem in self-driving cars) ▶ Bias in AI: How do we address biases in datasets and algorithms? ▶ Applications of AI and Ethics: Privacy (e.g., surveillance systems); Autonomous weapons; Mental health impacts of social media algorithms.

Aesthetics examples

▶ What is beauty? Is it subjective or objective? ▶ What makes something a work of art? ▶ How do we interpret and appreciate artistic expressions?

Aesthetics connection to AI

▶ Generative models: what is “creativity” when outputs are learned from data? ▶ Authorship and value: who (if anyone) deserves credit, and what is the work worth? ▶ Ethics around art: appropriation, and intellectual property in training data.

Philosophy of Language examples

▶ How does language refer to the world? ▶ What is meaning, and how is it conveyed? ▶ What is truth? ▶ Is there an ideal or universal language?

Philosophy of Language connection to AI

▶ How do we teach machines to understand and generate human language? ▶ How do AI systems link abstract symbols to real-world concepts? ▶ Context and ambiguity: why “plausible text” can still fail at meaning.

Logic examples

▶ What is valid reasoning? ▶ Can logic serve as the foundation for mathematics, science, and language? ▶ What are the limits of computability and provability?

Logic connection to AI

▶ Symbolic AI (e.g., Prolog programming for logical inference) ▶ Algorithms and Formal Languages ▶ What problems can AI solve, and what lies beyond its reach?