Why data contamination in large language models makes some AI benchmarks feel more like an open-book exam than a real test of intelligence—and what builders should do about it.