Persuasion–Accuracy Tradeoff in LLMs

A new arXiv study finds model scale boosts persuasive impact by roughly 1.6 percentage points per order of magnitude, with post‑training adding about 3.5 points. But increased persuasion correlates with reduced factual accuracy, implying optimization shifts models toward influence over truth. — This forces AI policy and evaluation to weigh manipulation risk against reliability, not just chase larger or more persuasive systems.