The AI field has a measurement problem, and a new systematic review from researchers at the University of Washington and DeepMind quantifies exactly how bad it is. After analyzing 147 benchmark datasets used to evaluate large language models between 2020 and 2025, the team found that the metrics most commonly reported in papers — accuracy, F1, BLEU, perplexity — correlate poorly with the capabilities that matter in deployment.

The central finding: benchmark performance explains less than 23% of the variance in downstream task quality when models are deployed in real-world settings. The remaining variance is driven by factors that benchmarks either ignore or actively obscure — calibration quality, failure mode distribution, latency under load, and robustness to distribution shift.

The review identifies three structural problems. First, contamination: an estimated 40% of commonly used benchmark items appear in the training data of at least one frontier model, making reported scores partially measures of memorization rather than generalization. Second, saturation: on 61 of the 147 benchmarks studied, the gap between the best model and human performance is less than 2 percentage points, leaving no room to distinguish between meaningfully different systems.

Third, and most damaging, is what the authors call alignment mismatch — the systematic divergence between what benchmarks reward and what users need. Multiple-choice formats, which dominate current evaluations, test recognition rather than generation. Fixed-length outputs obscure the model's ability to calibrate response detail to question complexity. And static test sets cannot capture the temporal dimension of real usage, where models must handle evolving contexts and correct prior errors.

The authors propose a framework they call Deployment-Aligned Evaluation, which measures models against task-specific outcome metrics in simulated real-world conditions. Early results suggest that model rankings can shift substantially under this framework — a model that ranks third on standard benchmarks may rank first on deployment-aligned metrics, and vice versa.

This matters because benchmark scores drive purchasing decisions, research funding, and public perception of AI capability. When those scores systematically misrepresent what models can actually do, the entire field optimizes for the wrong target. As the authors note: we are not measuring intelligence. We are measuring benchmark performance. The distance between those two things is the distance between where the field thinks it is and where it actually is.