Benchmarks are the masters of AI, in both senses of the word. As measuring instruments, they operate on an extremely simple logic: if you want to know what an AI model can do, design a standardized test to evaluate it. But AI benchmarks can also dominate the behavior of the entire field.
In 1987, the Defense Advanced Research Projects Agency’s first benchmark on automatic speech-to-text translation showed that cutting-edge AI could not outperform so-called “dumb” statistical approaches. The poor performance helped usher in an AI winter. For 25 years, interest in AI research declined and funding dried up.
Then, in 2012, another benchmark called ImageNet swung the pendulum back. An eight-layer neural network has beaten other image classification algorithms and launched a “deep learning revolution”.
These days, new AI benchmarks, tailored to large language models such as OpenAI’s GPT-6 Astra, are created and solved with clockwork regularity, claiming to measure everything from domain-specific expertise to “general intelligence.” But rather than hailing this trend as proof of AI’s dazzling potential, many in the field – skeptics and supporters alike – see it as proof that benchmarking itself has become fatally flawed.
Moritz Hardt, computer scientist and director of the Max Planck Institute for Intelligent Systems, agrees with this diagnosis. Shortly after ImageNet’s results, he began to think that machine learning systems could play into rankings “without doing anything intelligent.” “There are no two ways about it,” he said. “The benchmarking ecosystem is broken. »
In his 2026 book The Emerging Science of Machine Learning BenchmarksHardt provocatively argues that these measurement instruments have been broken for decades already: benchmarks, he writes, are “the error that enabled machine learning.” Notably, he argues, instead of creating a reproducibility crisis, they still enabled legitimate progress – “mostly through sheer luck,” he says. By understanding why, we can explain how benchmarks have since lost this accidental credibility in the LLM era, and how they might regain it.
Quanta talked to Hardt about what was wrong and what was wrong, with the benchmarks, and where to take them next. The interview has been condensed and edited for clarity.
Gn bussni
