Research indicates that as AI models tackle more complex tasks, failures are increasingly characterized by incoherence rather than systematic misalignment. The study identifies errors in frontier reasoning models as being composed of bias and variance components, with incoherence becoming more prevalent as reasoning lengthens.