CoBe: A Benchmark For Conversational Counterfactual Text Editing

Current AI systems struggle with causal reasoning in safety-critical domains: Even frontier language models achieved only an average score on CoBe's counterfactual editing task, suggesting that state-of-the-art AI cannot reliably distinguish between what should change versus what should remain fixed when reasoning about hypothetical scenarios—a critical capability for applications in healthcare, law, and other high-stakes fields.

AI models often confuse correlation with causation, creating plausible but incorrect outputs: The paper demonstrates that models frequently produce "associational" responses rather than true counterfactual reasoning. In a medical example, when asked what would happen if surgery weren't recommended, a model might incorrectly make symptoms less severe simply because milder cases statistically correlate with no surgery, rather than actually reasoning through the causal chain of events.

Existing datasets don't adequately test for causal understanding: Prior counterfactual datasets fail to clearly separate genuine causal reasoning from statistical association, meaning AI systems could appear more capable than they actually are. CoBe addresses this by explicitly testing zero-shot counterfactual editing across 2,000 scenarios in healthcare, science, business, and everyday situations.

This capability gap has real safety implications: The inability to properly reason about counterfactuals—understanding which consequences follow from a hypothetical change—poses risks in domains like healthcare and law where accurate causal reasoning is essential for responsible decision-making and advice.