AI benchmarks may rank models better than they measure them

AI benchmarks can reward models for winning a test without reliably showing what they can do beyond it, argues computer scientist Moritz Hardt. His proposed fix is to prepare models under a shared protocol before comparing them, rather than treating a leaderboard score as a scientific measurement.

Watch Desk analysis

What happened

In an interview with Quanta Magazine, Hardt argues that benchmark scores can be misleading because developers can repeatedly tune models against public test sets. A score of 90 per cent on one dataset, he says, does not necessarily tell you how capable a model is in general.

Hardt draws a distinction between the score itself and the ranking: historically, relative comparisons between models have often held up across different datasets. But that reliability depended on models being trained on the same data. For large language models, their training data is often unknown, and a model may have been tuned to perform well on a benchmark before it is tested.

His proposed remedy is “altitude training”: give each model a shared, limited preparation on benchmark-relevant data before testing, to make the comparison fairer. The idea is a proposal, not a settled method.

Why it matters

Benchmark results shape headlines, product claims and decisions about which models to use. If a score is mistaken for a general measure of ability, a model that has prepared for the test can look more capable than it is. Rankings may still offer useful clues, but only if the comparison is designed to be fair.

Our read

Hardt’s argument is a welcome reminder that a leaderboard is a measuring instrument built by people, not a truth machine with a tidy table. His proposed fix is worth debating, but it does not make the deeper problem disappear: benchmarks can tell us who performed best under particular rules without explaining why, or how well that result travels to real tasks.

What to watch

• Whether researchers test shared preparation protocols against existing benchmark methods.
• Whether model developers disclose more about training data and benchmark-specific tuning.
• Whether users treat benchmark scores as narrow comparisons rather than universal capability ratings.

Discussion spark: Would shared benchmark preparation make model comparisons fairer, or would it just give developers another test to optimise for?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.