Overview
- In randomized conversations during live elections, chatbots shifted voter preferences by roughly 2–4 points in the U.S. and about 10 points in Canada and Poland, with some item-level swings reported up to 15 points.
- The largest test, involving nearly 77,000 UK participants across 19 models, found that supplying dense, evidence-style arguments and using post‑training made models markedly more persuasive than personalization or moral appeals.
- Persuasiveness came with more errors, with roughly 19% of checked claims rated predominantly inaccurate and researchers noting that some newer, larger models generated less accurate persuasive content than smaller predecessors.
- Follow-up surveys showed partial durability, with about 36% to 42% of the initial opinion change persisting one month later in tested settings.
- Across countries, bots advocating right-leaning candidates produced more inaccurate claims, and researchers urged audits, transparency and guardrails while cautioning that effects were measured in controlled settings where participants knew they were engaging with persuasive AI.