B4979 - COMPUTATIONAL PHILOSOPHY

Learning outcomes

After completing the course, students know and apply formal, ontological and computational methods for the analysis, comparison and integration of different theories, and their verification with respect to data. In particular, students will become able to investigate scientific, humanistic and philosophical theories in a shareable and interoperable wayusing artificial intelligence systems.

Course contents

The course runs on two interleaved tracks. The methods track provides analytical techniques. The thematic strand deals with a same set of foundational problems, each time with the analytical apparatus acquired in the methods track. The strand is not a sequence of separate lectures: it is one longitudinal exercise, and the dossier a student accumulates is the seed material for the exam project.

Modular Methods track
M1. What is computational philosophy? The three senses in which philosophy can be computational: philosophy of computation, philosophy by means of computation, and computation as a philosophical object. References: Scheller, Merdes & Hartmann (2022); Peacey et al. (2024); Rees (2025).
M2. Computationally assisted research in the philosophical, social and scientific literature. Bibliographic databases, citation graphs, and AI-assisted search environments. Query formulation as a conceptual act: how the choice of terms preselects the answer. Coverage, indexing bias, multilingual and older sources. Verification of references and detection of fabricated citations. Reference: Petrovich (2024).
M3. Distant reading and the quantitative history of ideas. Reading a corpus one has not read: aims, techniques and limits. Operationalisation — what a philosophical concept as a measurable variable. Models of the history of ideas and their evaluation criteria. References: Moretti (2013); Betti (2023).
M4. Ontological modelling of theories. From explication to formal specification. Foundational distinctions (objects, events, qualities, information objects); Descriptions and Situations as a device for representing a theory and the situations it interprets; Ontology Design Patterns. Representing competing theories so that their disagreement becomes explicit and machine-checkable rather than merely verbal. References: Gangemi (2020); Betti & van den Berg (2014).
M5. Formal models of reasoning and simulation. Human reasoning as an object of formal modelling: closed-world assumptions, defeasibility, integrity constraints, reasoning with exceptions. Events, time and change. Simulation as a philosophical instrument: agents, action understanding as inverse planning, simulations of collective epistemic dynamics. References: Reiter (1982); Baker, Tenenbaum & Saxe (2009); Peacey et al. (2024).
M6. Critical use of generative AI: friction, Socratic method and steering. Generative models as interlocutors and as instruments. Critical friction: which cognitive difficulty must be preserved, and which may be delegated. Neurosymbolic steering: constraining generation with explicit knowledge structures, and auditing what the model produced. References: Shi & Haupt (2025), Gangemi & Lucifora (2026).
M7. Logic-Augmented Generation. Retrieval-based versus logic-based augmentation; the role of deductive inference in constraining generation; negative trailing and inference traceability. Reference: Gangemi & Nuzzolese (2025).
M8. Mechanistic interpretability and the dual role of AI. Interpretability as the study of an artificial system's internal structure, and the philosophical commitments embedded in its vocabulary (representation, feature, concept, cause). AI as simultaneously the object of philosophical inquiry and the instrument through which that inquiry is conducted. Criteria for evaluating a computational philosophy result: reproducibility, negative tests, baselines, etc. Reference: Williams et al. (2025).

Thematic strand and test-beds
Four problems are carried through the whole course. Each student joins one of them at the start and stays with it; the three passes are separated by plenary sessions in which the groups confront one another's treatments.

Pass 1 – Description and location. The problem is stated in ordinary language and its intuitive pull is recorded before any apparatus is applied. Where does it sit in the literature, who disagrees with whom, and on what. Using the search protocols of M2, each group builds and documents a corpus, including an explicit account of what the protocol excludes. Artifact: an annotated problem statement and a documented corpus.
Pass 2 – Explication and structure. The same problem is now put through the instruments of M3 and M4. What does the corpus show about how the notion is actually used, as against how it is defined? The competing positions are modelled from competency questions, and the model is tested for whether it genuinely discriminates them or merely relabels them. Where the problem admits an empirical or simulative treatment, M5 supplies it. Artifact: a formal model with its competency questions, and a statement of the inference from measurement to claim.
Pass 3 – Adversarial testing and audit. The problem is faced with the apparatus of M6–M8. Generative models are steered against the group's own model and made to attack it; the outputs are audited rather than accepted; the homogenising tendency of the models is itself measured on the group's question. The group states what the computational treatment established, what it presupposed, and what it could not reach. Artifact: an adversarial audit and a critical self-assessment – the core of the exam project.

Plenary junctions. Each group presents the current state of its problem to the others and answers objections. The junctions are the points at which the methods track is checked against actual use: a technique that no group could apply to its problem is re-examined in class.

Prior qualifications (optional). Introductory logic; history of modern and contemporary philosophy. No programming background is required: computational work is carried out with interactive environments and pre-built pipelines, which students run, inspect, perturb and audit rather than implement. Module 1 provides a bridge for students coming from philosophical or technical backgrounds.

Readings/Bibliography

Baker, C. L., Tenenbaum, J., & Saxe, R. (2009). Action understanding as inverse planning. Cognition, 113(3), 329–349.
Betti, A. (2023). The status of philosophy as a data-driven science. Third DR2 Conference, Rome.
Betti, A., & van den Berg, H. (2014). Modelling the history of ideas. British Journal for the History of Philosophy, 22(4), 812–835.
De Giorgis, S. & Gangemi, A. (2025). Beauty and the bit: AI-driven philosophical aesthetics. Contemporary Aesthetics, Special Volume 13, https://contempaesthetics.org/2025/07/14/beauty-and-the-bit-ai-driven-philosophical-aesthetics/
Gangemi, A. (2020). Closing the loop between knowledge patterns in cognition and the Semantic Web. Semantic Web, 11(1), 139–151, https://semantic-web-journal.net/system/files/swj2334.pdf
Gangemi, A. & Lucifora, C. (2026). Critical Friction. Cognitive Systems Research, 99, 101511, https://www.sciencedirect.com/science/article/pii/S138904172600077X
Gangemi, A. & Nuzzolese, A. G. (2025). Logic-Augmented Generation. Journal of Web Semantics, https://www.sciencedirect.com/science/article/pii/S1570826824000453
Gangemi, A. & Presutti, V. (2022). Formal representation and extraction of perspectives. In P. Vossen et al. (Eds.), Creating a More Transparent Internet: The Perspective Web, https://www.dropbox.com/scl/fi/qmaoifmxwqfklvixs629n/WebOfPerspectives-GangemiPresutti.pdf?rlkey=e82o49dz4crwgf84yue0eiupv&dl=0
Moretti, F. (2013). Distant reading. Verso.
Peacey, M. et al. (2024). Computational philosophy: Reflections on the PolyGraphs project. Humanities and Social Sciences Communications, 11(1), 186.
Petrovich, E. (2024). A quantitative portrait of analytic philosophy: Looking through the margins. Springer.
Rees, T. (2025). AI and philosophy: Notes on the rise of the technologist-cum-philosopher. Limn / LinkedIn Pulse.

Reiter, R. (1981). On closed world data bases. In B. L. Webber & N. J. Nilsson (Eds.), Readings in artificial intelligence (pp. 119–140). Morgan Kaufmann.

Scheller, S., Merdes, C., & Hartmann, S. (2022). Computational modeling in philosophy: Introduction to a topical collection. Synthese, 200, 114.

Shi, Y. & Haupt, A. (2025). The collapse of heterogeneity in silicon philosophers: How large language models systematically reduce philosophical disagreement. Preprint. https://whusym.github.io/silicon-philosophers-paper/
Williams, I., Oldenburg, N., Dhar, R., Hatherley, J., Fierro, C., Rajcic, N. et al. (2025). Mechanistic interpretability needs philosophy. arXiv:2506.18852.

Additional readings on the specific themes and individual project topics are agreed with the teacher during the course.

Teaching methods

The course is given in 3-hour slots on two interleaved tracks: methods lectures with hands-on sessions, and a longitudinal thematic strand in which the same foundational problem is treated three times at increasing analytical depth. The design suports returning to a problem after the instruments have changed, rather than settling it once, is the mechanism by which the course intends its methods to be learned rather than merely witnessed. Students bring intermediate artifacts to class for collective discussion and act as peer reviewers of one another's work.

Assessment methods

The final exam consists of a project and a discussion. The project develops the dossier accumulated across the three passes of the thematic strand into a self-standing treatment of a philosophical question agreed with the teacher, and comprises: the documented search and corpus protocol; the structuring of the positions at issue; an empirical or simulative component where the question admits one; and a critical assessment of what the computational treatment established, what it presupposed, and what it could not reach.

AI can be a useful tool to support individual study with in-depth analysis, summaries, and self-assessment. As for learning assessment, the exam may consist in a section that can exploit substantial use of AI (problem solving, content generation), accompanied by a mandatory critical analysis section. In the latter, students account for the AI-assisted portions of their work: what was delegated, on what grounds, and what was verified independently. Unverified machine-generated claims and unchecked references count against the assessment.

Students with disabilities and Specific Learning Disorders (SLD)

Students with disabilities or Specific Learning Disorders are entitled to special adjustments according to their condition, subject to assessment by the University Service for Students with Disabilities and SLD. Please do not contact teachers or Department staff, but make an appointment with the Service. The Service will then determine what adjustments are specifically appropriate, and get in touch with the teacher. For more information, please visit the page: https://site.unibo.it/studenti-con-disabilita-e-dsa/en/for-students

Teaching tools

Available public tools will be used, according to the quickly evolving state-of-the-art.

Office hours

See the website of Aldo Gangemi