A hallucination is a convincing but incorrect model response. In causal analysis, polished explanations and compiling code can still encode unsupported formulas or omit assumptions that determine whether an answer is valid. section 13.1.7 section 13.4
The practical danger is confusing fluent prediction with causal identification. A model may reproduce a plausible relationship without possessing, using, or exposing the level of information required to answer a causal query reliably. section 13.2.4 section 13.4
Probability of necessity
A generated answer supplied a well-formatted formula and Python implementation for probability of necessity, but the formula was valid only under monotonicity, absence of confounders, and equality of treatment and outcome probabilities. Without those identifying assumptions, the polished answer was a hallucination. section 13.1.7 section 13.2.4
Treat generated causal answers as hypotheses to audit:
- Check whether the query is identified with the information available, rather than accepting a fluent derivation. section 13.2.4
- Make assumptions explicit and use causal expertise to detect when omitted conditions threaten analysis quality. section 13.1.7
- Use LLMs for causal knowledge, DAG construction, formalization, and code support—not as a replacement for causal validation. section 13.1.7 section 13.4
