AI Benchmarks Miss Factual Truth

Ademola Balogun argues that popular AI benchmarks like MMLU and HumanEval measure reasoning and coding ability, not factual truth. Models scoring 95% on MMLU still invent citations. RAGTruth, an 18,000-response dataset, tracks hallucinations at the word level. Google DeepMind's FACTS Grounding benchmark targets factual reliability instead of benchmark scores.