The article argues against using large language models to infer causal structures, warning that such approaches risk confusing textual associations with genuine causal evidence. Instead, it proposes that agents should only assist the workflow by inspecting data and explaining assumptions, while leaving causal claims grounded in formal algorithms and diagnostics.
- Agents are restricted from supplying edges, orientations, priors, constraints, or causal conclusions to prevent hallucinated mechanisms from becoming evidence.
- The principle is instantiated in causal-learn+, an online platform coordinating data analysis, preprocessing, method recommendation, and interpretation within the causal-learn ecosystem.
- A case study on Big Five personality data demonstrates a pipeline that avoids turning language-model unreliability into causal evidence.
This approach ensures that causal discovery remains reliable by separating agent-assisted workflow coordination from the rigorous algorithmic grounding required for valid causal claims.