CoAX: Cognitive-Oriented Attribution eXplanation User Model of Human Understandi

CoAX — Cognitive-Oriented Attribution eXplanation — landed on ArXiv on May 1st, and it does something the explainable AI field has needed someone to do for a while: it turns the microscope around. For years, XAI research has operated on an implicit assumption that a good explanation, technically defined, is a good explanation full stop. Build the saliency map, generate the counterfactual, compute the SHAP values — and somewhere on the receiving end, a human being will understand. CoAX challenges exactly that gap. The paper proposes a user model — a formal account of how humans actually process, interpret, and ultimately trust or distrust AI explanations — grounded not in what engineers find elegant but in what cognitive science tells us about how understanding works. That is a different starting point, and it leads somewhere different. Rather than asking "did we generate an accurate attribution," CoAX asks "did anything useful land in the person's mind" — which turns out to be a much harder question with a much messier answer, one that depends on prior knowledge, cognitive load, the framing of the explanation, and a dozen other variables that feature attribution scores cheerfully ignore.

If confirmed, here is what this means. The XAI field may be building to the wrong specification. Every benchmark, every user study designed around proxy metrics of comprehension, every deployment of LIME or SHAP in a regulated industry setting — all of it sits on an assumption about human cognition that CoAX suggests has never been rigorously validated. The practical implications reach into healthcare AI, where clinicians receive model explanations that may feel interpretable without actually shifting their understanding in meaningful ways, and into AI regulation, where the EU AI Act and similar frameworks demand explainability without defining it in cognitive terms. If there is a systematic mismatch between what XAI systems produce and what human minds can absorb, the entire compliance apparatus built on explainability requirements becomes structurally unreliable. Second-order, this work opens a door for cognitive scientists to become necessary collaborators in AI development — not consultants brought in late, but architects of explanation systems from the start.

Watch for peer review uptake or citations from established XAI researchers, which would signal the field taking the cognitive modelling argument seriously. Any replication attempt applying the CoAX framework to an existing XAI benchmark would be the moment this moves from provocation to paradigm shift.

NewsHive monitors these sources continuously. All signal titles above link to the original reporting.

Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →