Essay topics for Philosophy and Theory of AI

Click here to get back to the course webpage.

Here I collect some possible topics. This is still in progress: I might add more topics and update the present ones.

Comments

Please take a look at the file formalities.pdf for requirements on the essay (word limit, format, etc.), including the grading criteria to get a clear idea of what a good essay is expected to look like.
As also described there, the topic below are just meant as first. It is part of your task of writing an essay to also cover the following.

• Argue for the importance of the chosen topic and identify a clear research question within the topic. The research question should be wide enough to capture an important aspect of the topic, but narrow enough to be answered in an essay. Ideally, the research question poses a yes/no question. Locate your contribution within related work.
• Your answer to the research question should involve philosophical or mathematical theory that we covered in class or that is closely related to the course.
• You should also describe an experiment to empirically test your answer to the research question. The description should be detailed in enough for a programmer to code it (which neural network architecture, which hyperparameters, which training data and how to get them, etc.; cf. the ‘machine learning pipeline’).
• Ideally, you also implement (a toy-version of) your experiment and describe the empirical results. However, this is not necessary, since coding is not required for the course. If you do not implement your experiment, then describe which results you would expect (e.g., by referring to papers with similar experiments).

Topics

• We discussed the question of holistic vs elementwise representation in the Phil Science section (Freiesleben, et al. 2022 “Scientific Inference With Interpretable Machine Learning”). A closely related distinction - reductionism vs holism - is discussed by Abramsky (2022, “Nothing will come from everything”), specifically mentioning the case of AI on page 544. In light of these ideas, discuss to what extent a holistic representation of neural network behavior is compatible with a compositional approach to interpretable machine learning. (You might also refer to a classic of modern philosophy, Quine’s ‘Two Dogmas of Empiricism’, as a critique of a too reductionist approach.)

• In Beigang (2023), counterfactual fairness of a prediction model was formalized using the counterfactual provided by causal models. Could it be formulated also using the counterfactuals from philosophy (the Lewis-Stalnaker counterfactual) or from explainable AI (see Wachter 2018 “Counterfactual explanations without opening the black box”)? For this, one could explore the variation semantics framework of Hudetz and Crawford (2022, “Variation semantics: when counterfactuals in explanations of algorithmic decisions are true”). What is the practical relevance of such results?

• In classical philosophy of AI (from the 1980s and 1990s), Gödel’s Incompleteness Theorem played a crucial role for arguing for impossibilities for (symbolic) AI. (See here for a brief overview.) Discuss and compare this to modern impossiblity result for machine learning like that of Colbrook et al. (2022, “The difficulty of computing stable and accurate neural networks”)?

• In an ideal case of interpretable AI, one can show how a neural network - that was trained to predict a real-world phenomenon - has learned a useful representation of this phenomenon from which it derives its predictions. Discuss how achievable this ideal case is. Specifically, does it even make sense to say that a neural network has ‘ representational capacities that are akin to mental representations’ (Freiesleben & Grote 2023, footnote 4)? (For a quick overview of `representation’ in AI, see here and also see the reading Millière & Buckner 2024 part 2, especially section 2.4.)

• One way to state the interpretability problem of neural networks (that I find particularly useful) is in terms of micro-level and macro-level. We completely understand neural networks at the micro-level: how they propagate the activation of the input neurons to the output neurons and how they update their weights during learning. The problem is understanding them at the macro-level: to explain in human-understandable terms this process that goes on at the micro-level. An analogy is provided by thermodynamics: we have a complete understanding of the billions of gas particles in a box (how their position and momentum changes over time), and thermodynamics also provides an understanding of this micro-level process in the macro-level terms of temperature and pressure. What can be said about this problem of relating the micro-level to the macro-level from a general philosophical theory of explanatory levels as, e.g., presented in this recent book and, in particular, in the first paper in there by List (which is also here)?

Example

As an example of a paper that covers the required steps, you can consider Geiger et al. (2025, “Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability”). Note, though, that it is much more comprehensive than what would be required for an essay. So this is just for illustration, not to be followed very closely. Very briefly, the idea is this:

• topic: mechanistic interpretability
• research question: can causal abstraction formalize mechanistic interpretability?
• related work: literature on mechanistic interpretability
• philosophical/mathematical theory: causal explanation à la Pearl and specifically the notion of causal abstraction
• experiments: a companion jupyter notebook for a toy example as well as many references to existing results in the mechanistic interpretability literature.