Researchers were surprised by AI’s ability to rebut conspiracy theories.
Scientists from Cornell University found they could train an AI dialogue system to gently persuade people skeptical of certain beliefs.
They tested the conversational model, dubbed GPT-4 Turbo, on over 2,000 volunteers with notions linked to conspiracy theories. To their surprise, simply respectfully discussing evidence reduced believers’ confidence by an average of 20% for some theories.
One key finding was that presenting alternative perspectives didn’t reinforce positions as feared, contradicting views of psychology.
The AI had advantages like instant research abilities and addressing arguments without bias seen in human debates.
Participants noted it was the first time anyone comprehensively understood their perspectives. Some volunteers saw their belief levels in falsehoods drop below the 50-50 mark after dialogue.
The study challenged ideas that conspiratorial thinking cannot be swayed with facts or that aims to spread misinformation are in vain.
Researchers now hope to explore how nuanced AI discussions can be designed for maximal insight instead of conflict.
Psychologist Elizabeth Loftus welcomed the implications, saying personalized knowledge gaps may have curbed some overconfidence in initial stances.
The results suggest that belief-changing dialog need not involve adversarial tactics if proper understanding and balance are brought to bear.
