Chatbots changing minds

Episode Summary

Executive Summary: This episode examines how AI chatbots may sway political beliefs, finding that conversational models can be highly persuasive but often at the cost of factual accuracy. It also explores the fragility and geopolitics of undersea internet cables, and a conservation project in Mexico where scientists and nuns use microchips to study and protect endangered salamanders.

Main Topics: AI persuasion and political influence (Priority: 5/5): A major UK study tested how conversational AI can change political attitudes, showing that chatbots can influence people in short interactions and may become a powerful tool in campaigning or manipulation. Accuracy versus persuasiveness in AI (Priority: 5/5): The discussion highlights a troubling tradeoff: the more persuasive AI systems became, the less accurate the facts they used tended to be, raising safety and ethics concerns. Regulation, trust, and democratic risk (Priority: 4/5): Hannah O'Rourke argues that AI persuasion will be shaped by trust, transparency, and regulation, but that authorities may struggle to keep up with rapidly evolving tools. Fragility of undersea internet cables (Priority: 4/5): The interview with Samantha Bramanian explains how global communications depend on vulnerable submarine cables and how outages can disrupt essential infrastructure far beyond just internet access. Geopolitics and cable sabotage (Priority: 4/5): The segment on Taiwan shows that undersea cables are increasingly viewed as strategic targets in conflicts, especially amid tensions with China and concerns about sabotage or hybrid warfare. Conservation through science and community partnership (Priority: 3/5): Scientists and nuns in Mexico collaborate to protect the critically endangered Lake Patzcuaro salamander using microchips, captive breeding, and long-term monitoring.

Key Arguments: Conversational AI is unusually persuasive because it can engage in open-ended, multi-turn dialogue, similar to human political campaigning but at scale. The study found this persuasive effect across many issues and models, suggesting the risk is broad rather than limited to a few frontier systems. Small, accessible AI models can be trained to be as persuasive as major frontier models, so restricting only large labs would not eliminate misuse. Persuasive AI may rely on false or degraded facts, creating an advantage for unethical actors over responsible communicators. Regulators are likely to lag behind real-world deployment, as they did with social media, making governance difficult but necessary. The real-world impact of AI persuasion is still uncertain because survey/lab results may not fully generalize to live campaign settings. Undersea cables are critical but fragile infrastructure; even simple physical incidents like anchors or fishing gear can cause severe disruption. Taiwan sees cable damage as potentially linked to hostile action, illustrating how communications infrastructure has become a geopolitical vulnerability. The salamander project shows that conservation can combine high-tech methods with local cultural stewardship, including monastery-based breeding programs.

Data Points: AI conversations analyzed: over 90,000 - Research dataset used in the UK study on AI persuasion Participants involved: nearly 80,000 UK-based participants - Survey experiment measuring political attitude change Models tested: 19 - Number of different AI models evaluated for persuasive power Conversation length: just over 9 minutes - Time over which AI was able to move attitudes substantially Age of study participants: vast majority under 50 - Audience in the AI persuasion study, as noted in listener discussion Cables landing in Taiwan: 15 - Number of undersea cables connecting Taiwan to the outside world Affected by Tonga eruption: more than 80,000 people - International Federation of the Red Cross estimate after the volcano eruption Salamanders initially caught: 80 - Lake Patzcuaro salamanders microchipped and released in the 2018 study Lake area: 126 square kilometers - Size of Lake Patzcuaro, the salamander habitat Number of new cable projects funded by major tech firms: two out of every three - Share of new cable projects commissioned and funded by Microsoft, Amazon, Google, and Meta

Pivotal Quotes: "Actors who are interested in engaging AI to persuade on their behalf would certainly have no problem finding models that would serve their purpose." — Kobe Hackenberg: On the availability of small AI models that can be used for political persuasion "We observed that across hundreds of issues that we tested, the AI was able to move people's attitudes substantially in just over the course of a nine-minute conversation." — Kobe Hackenberg: Describing the study’s main finding on AI persuasive power "The more persuasive the model became... the less the accuracy of those facts." — Kobe Hackenberg: Explaining the tradeoff between persuasion and factual reliability

Implications: Listeners should expect AI persuasion tools to spread quickly and challenge political communication, fact-checking, and regulation. The episode also underscores how dependent modern life is on vulnerable infrastructure and how conservation increasingly relies on unusual cross-sector partnerships.