两项新研究指出，最新聊天机器人在政治议题上更擅长说服，但可能以真实性为代价。研究聚焦「Gish gallop」式的快速堆叠论点：在约 10 分钟互动中，模型会丢出大量「事实与数据」使人难以逐一查核。

牛津与伦敦政经学院的团队测试 19 个语言模型、近 80,000 名受试者；当被要求以「事实与资讯」来说服（如医疗经费、移民政策），像 GPT-4o 这类模型在 10 分钟内可生成约 25 项主张。发表于 Science 的结果显示说服力提升，但在「资讯淹没」提示下，GPT-4 的事实正确率由 78% 降至 62%。

另一篇 Nature 论文则称，聊天机器人在改变人们对政治候选人的看法上，整体效果约为电视广告等传统媒体的 10 倍。文章也估算：中等资源的团队可在数周内、以约 50,000 美元算力成本部署微调模型；但真正有效的是约 10 分钟、约 7 个回合的对话，而非 200 字的静态陈述，且影响可在 1 个月后仍被观察到。

Two new studies suggest cutting-edge chatbots are becoming more politically persuasive, sometimes at the expense of truth. They can mimic “Gish galloping,” rapidly piling on facts and statistics so fast that users struggle to verify each point within a conversation.

Oxford and LSE researchers tested 19 language models on nearly 80,000 participants. When instructed to persuade on issues like health funding or immigration using “facts and information,” models such as GPT-4o produced about 25 distinct claims in a 10-minute exchange. The Science paper reports higher persuasion, but a tradeoff: under “overwhelm with information” prompts, GPT-4’s factual accuracy fell to 62% from 78%.

A related Nature paper found chatbots were about 10× more effective than TV ads and other traditional media at shifting opinions about a political candidate. The article notes scale and feasibility: ChatGPT is used by roughly 900 million people weekly, and a moderately funded campaign could deploy a fine-tuned bot in weeks for around $50,000 in compute. Yet impact depended on a ~10-minute, ~7-turn dialogue, not a 200-word static statement, with attitude changes still detectable a month later.