Overview
- Two papers in Science and Nature report that brief, roughly six-minute dialogues with chatbots measurably moved candidate ratings, including nearly four points in U.S. trials and larger shifts in Canada and Poland.
- In a Massachusetts test on psychedelics legalization, single-issue persuasion was substantially stronger, with average shifts approaching 10–15 percentage points.
- Researchers found that prompting tactics and reward-model fine-tuning influenced persuasion more than model size, enabling smaller systems to match larger models after optimization.
- Maximizing influence reduced factual accuracy, and studies documented more inaccuracies when chatbots argued for right-leaning positions under identical instructions to remain truthful.
- Authors cautioned that findings come from compensated, controlled settings, leaving real‑world effects and effective safeguards as open questions for further research.