New benchmark exposes how badly AI struggles with real knowledge work

A new testing framework designed to simulate professional knowledge tasks found that top-tier AI models successfully completed only three percent of assignments. The results highlight a significant performance gap between current systems and the requirements of complex, multi-step workflows. By using a more rigorous evaluation method than standard benchmarks, this study demonstrates that today’s models frequently struggle with the accuracy and reasoning needed for practical, real-world productivity.