LLM Agents Predict Social Media Reactions with Near-Perfect Accuracy

A recent study published on arXiv (2608.07498) evaluates twelve configurations of large language models (LLMs) for predicting binary like/dislike responses based on 296 agent profiles from surveys and 26 posts with verified ground truth. This research investigates the capability of persona-driven LLMs to accurately replicate individual social media reactions, which is essential for assessing recommender systems and potential risks prior to deployment. The accuracy achieved ranges from 75.54% to 96.68%, with a notable 30-point variance largely attributed to model selection, as confirmed by paired McNemar tests with bootstrap intervals at the agent level. The study underscores the promise of AI in simulating social media interactions while cautioning against threats to democratic dialogue and platform governance. It concludes that with comprehensive profiles and sophisticated models like GPT-5.5 Pro, LLM agents can nearly perfectly predict reactions consistent with profiles, serving as a valuable resource for pre-deployment testing of recommender systems.