Lie to persuade

Fascinating article in Science on "The levers of political persuasion with conversational artificial intelligence" by a multidisciplinary team (Kobi Hackenburg, Ben Tappin, Luke Hewitt, Ed Saunders, Sid Black, Hause L., Catherine Fist, Helen Margetts, David Rand, Christopher Summerfield). Clever experiments, lots of caveats for sure, but within the context of their setup, some disturbing, albeit not entirely unexpected, findings.

➡️ They ran "three large-scale experiments with 76,977 total UK participants, we deployed 19 large language models (LLMs) to persuade on 707 political issues while varying these factors independently. [They] also analyzed more than 466,000 AI-generated claims, examining the relationship between persuasiveness and truthfulness."

And the last bit is the one that should worry us: "When AI systems are optimized for persuasion, they mayincreasingly deploy misleading or false information."

Which begs the question: are politicians post-trained chatbots?