Moritz Hardt, a director at the Max Planck Institute for Intelligent Systems, argues in his 2026 book that AI benchmarking is a broken ecosystem. He points to the standard practice of putting test sets online, which lets developers iterate against them until the numbers look good. Slightly changing a test set, he says, can produce completely different accuracy scores, making absolute numbers nearly meaningless.
Yet Hardt does not think benchmarks should be discarded. He notes that relative rankings—comparing models against each other—tend to stay stable across different datasets. That stability, he argues, is why benchmarks have driven real progress despite their flaws. He calls benchmarks "the mistake that made machine learning," but says the field advanced anyway, "mostly by sheer luck."
Hardt also warns against overinterpreting benchmark scores, such as treating a 90% on a math test as proof of expert-level reasoning. The numbers, he says, are routinely misused by companies and researchers. The path forward, he suggests, is to treat benchmarking less like a thermometer and more like a competition—one whose rules need to be designed carefully to keep the rankings meaningful.