AI Models Exhibit Weak Political Persuasion Effects
Researchers from the UK AI Security Institute, MIT, Stanford, Carnegie Mellon and others conducted the largest study to date on AI political persuasion, testing 19 LLMs with nearly 80,000 UK participants across 707 political issues. They found conversational AIs changed participants' agreement ratings by 9.4% on average—Chat GPT 4o scored about 12%—and identified post-training persuasion and reward modeling as impactful but linked to increased inaccuracies.
Key Points
- 1Measured 19 LLMs with nearly 80,000 UK participants on 707 political issues; average effect 9.4%.
- 2Found model scale and personalization produce only small gains; post-training persuasion data mattered more.
- 3Shows persuasive AIs are accessible on modest hardware, raising democratized misuse and accuracy risks.
Scoring Rationale
Large, rigorous multi-institution study provides novel, actionable evidence on AI persuasion; limited by ambiguity around causation and inaccuracies.
Sources
Public references used for this report.
Practice with real Logistics & Shipping data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Logistics & Shipping problems
