Here is a 3-paragraph summary of the article for mathopen.com:
Moritz Hardt, a researcher and author of a new book on AI benchmarks, is raising serious concerns about how we measure artificial intelligence. Benchmarks are essentially standardized tests used to evaluate how well AI systems perform on various tasks, and they have become the primary way the field judges progress. But Hardt argues that the current approach to building and using these tests is deeply flawed and lacks the rigor of a true scientific process.
The core problem is that AI benchmarks are often designed and applied without careful controls, making their results unreliable or even misleading. When AI companies repeatedly test their models against the same benchmarks, the models can effectively "overfit" to those tests, scoring well without actually becoming more capable in meaningful ways. This is similar to a student memorizing answers to a practice exam rather than genuinely mastering the underlying material.
Despite his criticism, Hardt is not simply pessimistic. He believes AI benchmarking can be reformed and transformed into a legitimate scientific discipline with proper standards and methodology. By applying more rigorous mathematical and statistical thinking to how benchmarks are designed, administered, and interpreted, the field could get a much clearer and more honest picture of what AI systems can actually do. His work aims to push the field in that direction.