AI's biggest blind spot isn't politics, it's your health (www.techradar.com)

HUMAINE, a large-scale study of over 40,000 anonymized UK and US conversations, found that age—not politics—best predicts which AI tools people prefer, and revealed a striking behavioral trend: health and wellbeing dominate user interactions. Nearly half of conversations centered on proactive wellness (fitness, nutrition), while many involved deeply personal mental-health and medical queries. That makes clear the practical reality of LLMs acting as informal therapists and health advisers, and exposes a critical mismatch between how models are evaluated and how they're actually used. The study argues current evaluation regimes—academic benchmarks for abstract skills and public “arenas” driven by crowd votes—miss crucial human-centered dimensions. Trust-and-safety ratings were noisy and inadequate for capturing high-stakes harms, so the authors and cited work (e.g., Stanford HAI) call for scenario-based, standardized tests (examples: CIP’s weval.org) that probe mental-health failures and long-term effects. The piece also warns metrics that optimize only for task completion risk mindless automation and workforce de-skilling; instead we should measure collaboration, learning, demographic consistency, and robustness. Practically, Gemini-2.5-Pro topped HUMAINE for cross-metric consistency, suggesting progress comes from broadly reliable systems rather than narrow high scores. The takeaway: build richer, population-aware evaluations to surface AI’s real-world blind spots—especially in health.