Can we extract causal belief networks from interview conversations?

Building generative agents that simulate human reasoning rather than merely produce plausible stances — a goal motivated by the Can language models simulate belief change in people? critique — requires modeling individual internal logic. The proposed pipeline is a three-step process from natural language to executable causal belief networks (CBNs).

Step 1: Extract causal motifs from QA responses. LLM-conducted semi-structured interviews adaptively elicit causal explanations in everyday language ("why do you support X?" "what does Y influence?"). Responses are annotated with concept nodes and directional relations. For example: Q: "How might surveillance affect public safety?" A: "It can reduce crime by aiding investigations with more transparency, which increases public safety." Motif: Transparency → Crime rate → Public safety. A second QA might add: Privacy ← Transparency → Crime rate.

Step 2: Compose a Causal Belief Network. The motifs are compiled into a belief graph representing the participant's reasoning. Nodes are concepts (fairness, safety, family needs); edges are directional causal relations with confidence and polarity scores derived from motif density or respondent emphasis.

Step 3: Simulate belief change via intervention. Apply a hypothetical intervention such as do(Transparency = high), reflecting a policy shift like increased camera accountability. Use belief propagation over the CBN to update downstream posteriors. Example: P(Privacy Concern) shifts from 0.7 to 0.3, and P(Opposition to Surveillance) shifts from 0.7 to 0.2.

The chain demonstrates that motif-based causal modeling can simulate how individuals update beliefs in response to policy changes, moving beyond static opinion snapshots. But the paper acknowledges open challenges: extracting CBNs from natural language remains hard due to ambiguity in concept identification, causal direction, polarity, and conceptual granularity. And causality alone cannot capture the full range of human reasoning — people also rely on associative, analogical, and emotional processes that resist strict symbolic modeling. The initial focus on causality is described as a strategic and computationally tractable starting point, not an endpoint.

The pipeline's value is that it makes belief simulation auditable. Unlike a prompted persona whose reasoning is inscrutable, a CBN exposes the structure of belief and supports formal analyses (intervention, counterfactuals, sensitivity to specific edges). Policy simulation requires this auditability because stakeholders must be able to ask why a simulated agent reached a stance — an auditability requirement Can AI agents learn people better from interviews than surveys? approaches with content richness but not with structural transparency.

Source: World Models