Dataset of 2k+ LLM chats with real voters released to build better decision evals
logangraham · x · 2026-11-04
A new dataset of 2k+ LLM conversations with real US voters about the 2026 Midterms reveals how people actually prompt AI about elections: mostly basic info requests, predictable multi-turn patterns, and wide model differences in recommending candidates. The goal is better real-world decisionmaking evals; related work explores moral-psychology-based eval design.
More from Safety
- Nvidia lost credibility in Washington by exaggerating China's AI chip capabilities, CFR fellow says — mchorowitz · 2026-11-04
- Anthropic's RSP update says it may 'imminently' hit AI-accelerated R&D threshold — thlarsen · 2026-11-04
- METR-Redwood red team found ~10% of OpenAI's internal Codex traffic went unmonitored — dl_weekly · 2026-11-04
- Perplexity wires its open Decider model into Numbat v0.3.0 for agent security monitoring — AravSrinivas · 2026-11-04
- Reminder: You're legally liable for your AI agents' actions — 'whoops' won't hold up in court — var_epsilon · 2026-11-04
- CrowdStrike unveils AI Kill Switch to give humans runtime control over AI agents — AccBalanced · 2026-11-04