Measuring What Matters: Construct Validity in Large Language Model Benchmarks

A very large review of AI benchmarks that reveals a worrying trend in their effectiveness and scientific rigor