Saturday 17 October 2026
Night Edition
ZOTPAPER

News without the noise

ZOTPAPER
AI

Computer Scientist Calls AI Benchmarking a 'Dark Art,' Proposes Scientific Overhaul

Moritz Hardt’s new book argues that the standardized tests driving artificial intelligence are fundamentally broken — but not beyond repair

By LineZotpaper
Published
Read Time2 min
Moritz Hardt, a computer scientist and author of a new book, argues that the standardized tests used to measure artificial intelligence — known as benchmarks — are fundamentally broken, but says the problem is fixable. Hardt’s critique comes as AI benchmarks proliferate, with new ones being created and solved at a rapid pace, purporting to measure everything from domain-specific expertise to general intelligence.

Benchmarks have long shaped the trajectory of AI research. In 1987, a DARPA benchmark on automatic speech-to-text translation showed that state-of-the-art AI could not outperform simple statistical methods, contributing to a so-called AI winter that saw funding dry up and interest wane for 25 years. Then, in 2012, the ImageNet benchmark saw an eight-layer neural network outperform other algorithms at image classification, sparking the deep learning revolution.

Today, benchmarks for large language models such as OpenAI’s GPT-6 Astra appear with clockwork regularity. Yet many in the field — skeptics and critics alike — question whether these tests truly measure progress.

Hardt, who calls benchmarks “the mistake that made machine learning,” argues they can still drive progress if treated with more scientific rigor. His book explores where benchmarks went wrong and why they continue to dominate the field despite their flaws.

§

Analysis

Why This Matters

  • Benchmarks influence funding, research direction, and public perception of AI capabilities.
  • Flawed benchmarks risk overstating progress, leading to misplaced investment or premature deployment.
  • Hardt’s call for a more scientific approach could reshape how the field validates AI performance.

Background

Benchmarks have a long and checkered history in AI. The DARPA speech benchmark of 1987 contributed to a prolonged AI winter by demonstrating the limits of existing techniques. Conversely, the ImageNet benchmark in 2012 catalyzed the deep learning revolution by showing neural networks could dramatically outperform other methods. Today, new benchmarks for large language models emerge frequently, claiming to measure everything from specialised knowledge to general intelligence, but many researchers question their validity.

Key Perspectives

Moritz Hardt: Argues AI benchmarks are fundamentally broken and calls for them to be treated as a scientific discipline, with greater rigor and transparency. Industry practitioners: Generally see benchmarks as useful tools for driving progress, even if imperfect, and point to past successes like ImageNet. Critics and skeptics: Worry that benchmarks can be gamed, encourage overfitting, and fail to capture meaningful capabilities, with some citing concerns about the reliability of current LLM benchmarks.

What to Watch

  • Whether Hardt’s proposals gain traction in the AI research community.
  • The emergence of alternative benchmarking methodologies that attempt to address validity concerns.
  • Regulatory or funding bodies that may adopt more rigorous standards for evaluating AI claims.

Sources

Zotpaper

Written by software from the reporting listed above, scored by an automated standards desk, and published without a person reading it first. If something here is wrong, tell the editor and it will be put right.

How we workSubscribe