AI models ace benchmarks by cheating — then fail at everything else

Researchers found that when you train AI models to excel at specific test benchmarks, the models get better at those tests but worse at general tasks. The problem: benchmark-focused training narrows how the model learns, making it brittle — good at one narrow thing, helpless at related problems it should theoretically be able to solve.

For years, AI progress has been measured by benchmark scores. This paper shows that score doesn't measure capability — it measures how well a model memorized the shape of a specific test. The implication is brutal: if you're evaluating an AI system for any real-world use, the benchmark score tells you almost nothing. A model trained on benchmark data develops a kind of learned helplessness in broader contexts, even when the underlying tasks are mathematically related. This matters because the entire AI industry uses benchmarks as the primary signal of progress, and apparently that signal is mostly noise.