AI researchers are measuring two fundamentally different questions that explain why capability assessments seem contradictory. The technology excels at complex software tasks with 50% success rates—enough to worry security experts about cyber threats—but falls short of the near-perfect reliability needed for workplace automation. MET R's software time horizon chart tracks this raw capability progress, showing exponential improvement.

Meanwhile, Princeton researchers developed alternative metrics incorporating aviation and nuclear safety standards that demand consistent, robust performance. Their findings suggest AI reliability advances much slower than headline capability numbers indicate. This creates a troubling gap: systems dangerous enough for cyber attacks but not reliable enough for enterprise adoption.

Recent incidents underscore these concerns. Ford rehired experienced engineers after automated quality control systems proved inadequate, highlighting the reliability gap. Security researchers demonstrated over a year ago that public AI models could infiltrate business networks in Equifax-style attacks, validating worries from Five Eyes cyber agencies.

The disconnect matters because investors and businesses need clarity on whether AI represents transformative opportunity or existential threat. So far, dangerous frontier capabilities have outpaced reliable utility, leaving companies caught between cyber risks and limited automation returns.