The Hidden Cost Behind AI's Accuracy Claims

Google's AI Overviews achieved 91% accuracy on the SimpleQA benchmark, but this headline masks a deeper crisis in AI search reliability. While the model improved between Gemini 2 and Gemini 3, the percentage of answers that lacked factual grounding from their cited sources actually increased from 37% to 56%, revealing a structural flaw in how AI summarization faithfully represents evidence. This divergence between accuracy metrics and actual trustworthiness exposes the false sense of security that current AI search interfaces create for users.