Lunchtime Talk - Aydin Mohseni 9/19/25

Value of the Information & Strength of the Argument

The talk provides a novel and significant contribution by challenging a widely cited result in causal inference. The argument is well-structured, building on formal definitions and a clear analogy. The speaker carefully delineates the scope of the reduction, acknowledging limitations and open questions. The use of de Finetti’s theorem as a conceptual bridge is illuminating and strengthens the argument. However, the presentation is dense and assumes familiarity with SCMs and Bayesian epistemology, which may limit accessibility. The argumentation is solid, but the proof sketch is not fully detailed in the talk, and some steps rely on assumptions that may be contested.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates high scientific rigor, referencing key works such as Pearl and Mackenzie (2018) and Bareinboim et al. (2022). The speaker clearly states the assumptions and scope of the reduction. The title accurately reflects the content, though it is generic. The talk does not include a public Q&A, but the speaker invites questions, indicating openness to scrutiny. The sources cited are appropriate and directly relevant to the argument.

The title accurately reflects the content: a lunchtime talk by Aydin Mohseni on a formal result in causation and probability.

8/10

The talk presents a formal proof of a reduction of causal claims to probabilistic ones within structural causal models, challenging a well-known irreducibility thesis. The argument is rigorous, builds on prior work, and is delivered by a philosopher with relevant expertise. However, the presentation is a research talk, not peer-reviewed, and some claims rely on assumptions that may be debated.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 02:00Introduction and acknowledgments of collaborators and teachers.
• 05:00Overview of the talk's aims: challenging the Pearl Causal Hierarchy and proposing a reduction.
• 10:00Explanation of the Pearl Causal Hierarchy and its three levels: associational, interventional, counterfactual.
• 15:00Formal definition of structural causal models and the languages L1 and L2.
• 20:00Example of irreducibility: gene as common cause vs. smoking causing cancer, indistinguishable by observation.
• 25:00Discussion of Bareinboim et al.'s proof of generic irreducibility and its interpretation.
• 30:00Key insight: interventions can be endogenized, making them part of the probabilistic model.
• 35:00Analogy with de Finetti's representation theorem for exchangeable sequences.
• 40:00Outline of the proposed representation theorem for causal learning and reasoning.
• 45:00Implications for decision theory, AI, and the philosophy of causation.

• The Book of Why: The New Science of Cause and Effect — Pearl and Mackenzie (2018) advance the Pearl Causal Hierarchy thesis.
• On Pearl's Hierarchy and the Foundations of Causal Inference — Bareinboim et al. (2022) claim to prove the irreducibility of causal claims to probabilistic ones.
• La prévision: ses lois logiques, ses sources subjectives — de Finetti's representation theorem for exchangeable sequences.

• Causality: Models, Reasoning, and Inference — Pearl's foundational work on causal models.
• The Book of Why — Presents the causal hierarchy and its implications.

• On Pearl's Hierarchy and the Foundations of Causal Inference — Bareinboim et al. argue for the irreducibility of causal to probabilistic claims, which the talk challenges.