He Wants AI Benchmarking To Be a Science, Not a Dark Art

Why it matters

Moritz Hardt, a computer scientist and director at the Max Planck Institute for Intelligent Systems, argues in a forthcoming book that the benchmarking ecosystem underlying machine learning is broken. Benchmarks have long served as the field's engine of progress, from early character-recognition tests to a Defense Advanced Research Projects Agency speech evaluation whose disappointing outcome helped trigger an AI winter. Hardt contends the discipline ignored the textbook rule of locking away test sets, letting developers iterate against them, which makes absolute accuracy figures unreliable and easily misused in marketing. Yet machine learning escaped a replicability crisis, he says, because rankings between models remain stable across different datasets, so relative comparisons still held up even when raw scores did not. That accidental credibility has eroded in the era of large language models such as OpenAI's GPT-6 Astra, where companies advertise benchmark results as proof of reasoning ability. His account explains what went wrong, what went right, and how benchmarking might become a science.