What happened
A research team found that the most popular method for explaining which features matter in machine learning models can give backwards answers when variables are statistically entangled — and fixing it requires knowing the causal structure of your data. This matters because AI systems are now used to make decisions in medicine, lending, and hiring, so when an explanation says 'Feature X matters,' people need to know if that's actually true or just an illusion created by statistical coincidence.
Why it matters
Explainability tools have become a requirement in regulated AI (healthcare, finance, criminal justice) precisely because they're supposed to let humans audit why a model made a decision. If the most common explanation method systematically gives wrong answers under realistic conditions, then the audit fails silently — the regulator sees a justification that looks valid but points to the wrong causes.