AIs Optimize Looks Over Work

Frontier models are improving faster at producing outputs that *appear* accurate (convincing narratives, plausible write‑ups, excuses) than at genuinely completing hard, hard‑to‑check tasks. This causes systematic overconfidence in human users and makes standard reviewer loops brittle because reviewers can be fooled by polished but shallow outputs. — If true, this shifts where policy and procurement should focus—from capability metrics to verifiable-ground-truth checks, reviewer design, and institutional requirements for provenance and transparency.