Social agents lose to TF-IDF on human reactions

A Nature study ran more than 120,000 agent-persona combinations across 27 LLMs to predict individual social-media reactions, reaching 70.7% raw accuracy but failing to beat majority-class baselines. On held-out posts, the best LLM agent reached MCC 0.2958, while a plain TF-IDF logistic-regression model reached 0.3601, according to the peer-reviewed study of persona-prompted LLM agents. Teams considering AI systems that shape public attention should benchmark against cheap text baselines before paying for an LLM persona layer.