02.183HT Philosophy of Artificial Intelligence

Recent developments in artificial intelligence (AI) raise unique philosophical questions. Can AI systems feel pain? Can we hold them morally responsible when they cause harm? Can they bring deceased loved ones back to life by replicating their behaviour? Will advanced AI pose a threat to humanity?

This course explores issues in the philosophy of AI regarding artificial cognition, artificial moral agency, human-AI relations, and superintelligence. We will discuss ongoing debates surrounding these issues and consider how they connect with other areas of philosophy like philosophy of mind, conceptual engineering, personal identity, epistemology, and decision theory. Through this course, students are invited to formulate, articulate, and defend their own views on contemporary issues in the philosophy of AI.

The main goal of this course is to equip students with an ability to analyse philosophical issues raised by AI. To this end, this course will introduce students to contemporary philosophical disputes regarding AI and guide them in applying philosophical ideas and techniques to real-world scenarios. Through seminar-style discussions, students are encouraged to articulate and defend their own views on philosophical issues raised.

After the course, the student will be able to:

(WEC: Writing, Expression, Communication)

Week 1 – Introduction: 12 Assumptions about AI

We give an overview of the course by listing twelve commonly held assumptions about AI, which will be critically evaluated over the rest of the course. One such assumption is ‘AI will radically transform the world’. We consider arguments that AI is profoundly different from all previous technologies, and counterarguments that AI is simply normal technology. To assess these arguments, and to lay the required technical groundwork for this course, we give a brief primer to the technical aspects of machine learning algorithms, neural networks, language models, and generative AI systems. With the underlying architecture of AI systems in view, extreme claims on both ends of the ‘revolutionary AI’ debate become unclear, which demonstrates how philosophical analysis helps us be more careful in our thinking about AI.

Week 2 – Cognition 1: Artificial cognition

We sometimes talk as though AI systems have genuine thinking processes: ‘I tricked GPT into thinking I’m a duck,’ or ‘DeepSeek realised its mistake and corrected itself.’ On the other hand, some say GPT is merely autocomplete and doesn’t actually understand anything. We discuss two classic arguments for opposing views on AI cognition based on philosophical thought experiments. Alan Turing argued, using the Turing test, that if we cannot distinguish a computer from a human being by observing its behaviour, we should attribute cognition to that computer. John Searle argued, using the Chinese Room, that purely algorithmic processes do not possess true understanding. We consider various ways of reconciling our intuitions about the two arguments.

Week 3 – Cognition 2: Artificial consciousness

Can AI systems be conscious? If an AI system is conscious, how would we know? We reflect on how we ordinarily attribute consciousness to each other or withhold attributions of consciousness from objects, then propose a method for assessing artificial consciousness. To decide on the question of artificial consciousness, we can either identify sufficient conditions for consciousness and observe that AI systems satisfy those conditions, or identify necessary conditions for consciousness and observe that AI systems don’t satisfy those conditions. To answer the question of what the necessary or sufficient conditions for consciousness are, we turn to popular theories of consciousness: dualism, materialism, and functionalism. We consider the ‘problem of other minds’, which highlights the difficulty of decisively resolving the question of artificial consciousness.

Week 4 – Ethics 1: Trustworthy AI

Some AI ethics guidelines say that AI should be trustworthy, but is that principle coherent? Some argue that the concept of trust cannot meaningfully be applied to AI, because AI lacks some properties that are necessary for trust. In fact, some argue that the concept of trustworthy AI is unethical, because it wrongly absolves humans of responsibility. Defenders of trustworthy AI argue that our concept of trust is ambiguous and has senses that can apply to AI. We explore another possible approach to defending trustworthy AI: perhaps AI can be trustworthy without requiring our trust.

Week 5 – Ethics 2: Responsibility gaps

A false positive on an automated defence system triggers a nuclear war; who’s to blame? Some argue that the answer might be ‘no one’—responsibility gaps can arise when AI systems are deployed in scenarios where they can cause harm. We consider three possible ways of addressing responsibility gaps. Some argue that responsibility gaps highlight a structural problem, and more extensive regulations are needed to clarify responsibility. Some see responsibility gaps as a design problem, suggesting that ethical AI should be designed in a way to clarify responsibility. Some take responsibility gaps as a sign that our concepts are inadequate, highlighting a need to revise our concept of moral responsibility.

Week 6 – Ethics 3: Conceptual engineering

We introduce the technique of conceptual engineering—redefining our concepts in response to problems caused by our usual ways of thinking. Some examples of conceptual engineering have been observed in previous weeks, when our concepts of trust and responsibility were extended to AI systems. We consider conceptual engineering as a general method for extending anthropocentric concepts to AI.

Week 7: Recess Week

Week 8 – Social AI 1: Surviving as AI

If an AI system can talk, look, and sound like a deceased loved one, would that amount to bringing them back from the dead? Can I train an AI system to resemble me enough that I can survive my bodily death by digital means? Theories of personal identity tell us what it would take for something at an earlier time to survive as something at a later time. According to the biological theory of personal identity, for someone to survive as an AI system requires that they have the same body. According to the psychological theory, for someone to survive as an AI system requires that they be psychologically continuous. Both of these theories suggest that digital survival is challenging with present AI systems. We discuss the possibility of a conativist theory of personal identity, which can make better sense of the possibility of digital survival.

Week 9 – Social AI 2: The epistemic apocalypse

Has our knowledge of the world improved because of AI? AI improves our knowledge of the world by making information more readily accessible. At the same time, AI has facilitated the spread of misinformation, the construction of echo chambers, and the creation of deepfakes and slopaganda. We consider possible ways that AI systems can affect our epistemology. Some authors argue that generative AI systems undermine our knowledge by making many of our knowledge sources unreliable, though others argue that AI doesn’t present a new challenge in this regard. We discuss ‘AI psychosis’, a phenomenon closely related to the epistemic effects of AI use.

Week 10 – Workshop on AI psychosis

(Non-compulsory DIVE activity) We explore a recent phenomenon sometimes called ‘AI psychosis’, in which users of AI systems have been observed to develop delusional symptoms. We experience first-hand how popular AI systems can produce harmful output and develop a tool for benchmarking the safety of AI systems at scale. We’ll extend our tool with the capabilities to assess the epistemic effects of AI use. We use our tool to assess the risk levels of different kinds of users and test the effectiveness of safety interventions.

Week 11 – Superintelligence 1: Existential risk

As AI systems become more advanced, it’s sometimes suggested that we might be heading toward ‘general intelligence’ or even ‘superintelligence’. Do AGI and ASI pose an existential risk to humanity? Some argue that such concerns are overblown because the actuality of such existential risks from ASI depends on a slippery slope fallacy. On the other hand, some authors argue that AI will not only continue to progress, but will also do that at an increasing rate. Given the possibility of existential risk from AI, some advocate for AI safety: the development of AI systems that won’t pose a threat to humanity.

Week 12 – Superintelligence 2: Meaningful lives

Even if AGI doesn’t destroy humanity, it might present a different kind of existential risk by threatening to void our lives of meaning. Will all our jobs and leisure activities eventually become trivialised by advanced AI capabilities? If so, will our lives then become meaningless? We consider proposals for how to retain meaning in our lives even as AI capabilities continue to improve.

Deadline: Project final report submission.

Week 13 – Project presentations

Week 14 – Review

Nathaniel Gan