当美国选民向ChatGPT或微软Copilot询问候选人信息时，他们得到的答案，可能来自一个自己从未听说过的小网站。这个网站叫Ballot Brief，收集了数千名候选人的资料。它的创建者本·托雷斯·埃兹里克（Ben Torres Ezrick）希望帮助选民更方便地了解竞选信息，也希望吸引一类特殊的访客——AI聊天机器人。
《华盛顿邮报》的一项测试显示，ChatGPT在22%的选举相关回答中引用了Ballot Brief，微软Copilot的这一比例达到75%。但这家网站上的候选人资料存在多处错误，其中一条错误还被Copilot重复给了提问者。
随着越来越多美国人向AI询问政治和选举信息，这项调查引出了一个问题：当聊天机器人依赖由AI生成的网站，错误会怎样传到选民面前？
一个小网站，成了AI的常用来源
埃兹里克曾在Google、Zoom和视频平台Rumble担任高级职务。他发现，选民想了解候选人时，往往需要花不少时间在不同网站之间查找，于是决定建立一个提供选举信息的一站式网站。
面对数千名候选人的资料，他选择让AI帮忙。他派出多组数字智能体在网上搜集信息，再将整理出的内容放进Ballot Brief，供人们查阅。
网站吸引到的访客远不止人类。埃兹里克称，他掌握的数据显示，在跨越9月和10月的30天里，微软的AI系统引用Ballot Brief的次数超过68万次。
《华盛顿邮报》的调查发现，Ballot Brief只是其中之一。包括它在内的三家新网站，经常出现在聊天机器人提供的投票信息来源中。三家网站的创始人均表示，他们的团队规模很小，几乎完全依赖AI收集和总结数据。
问题随之而来。民主与技术中心选举分析师伊莎贝尔·林泽（Isabel Linzer）认为，调查结果说明，一个人发布在AI生成网站上的信息，可能很容易就被聊天机器人采纳。她说，这可能影响选民对选举信息的信任。
错误被AI重复了一遍
《华盛顿邮报》在Ballot Brief上发现了多处错误。
该网站的一份资料称，俄亥俄州共和党参议员乔恩·赫斯特德（Jon Husted）支持过一项法案。但那项法案在他就任之前就已通过。Copilot在回答《华盛顿邮报》的问题时，重复了这一错误。
网站还将加利福尼亚州两名同叫罗伯特·加西亚（Robert Garcia）的竞选人合并进了同一份资料，另外几名众议院议员的委员会职务也被列错。
一些议员的资料则相当简略，缺少投票记录，网站还错误地标注称，他们似乎没有竞选网站。
埃兹里克表示，他努力确保信息准确。在《华盛顿邮报》向他展示调查结果后，他对网站的运作方式作出了多项调整。
他说，网站的目的是尽可能向选民提供更多信息，但不应成为他们获取信息的唯一来源。
一次覆盖3100个县的测试
为了了解聊天机器人如何回答选举问题，《华盛顿邮报》与数据公司Promptwatch合作，对OpenAI、Anthropic、Google、微软、亚马逊和Perplexity的AI产品进行了测试。
研究人员针对美国约3100个县，每个县选取一个地址，向聊天机器人提出同一个问题，请它根据该地址制作一份2026年11月选民指南，列出所有选举以及候选人对重大议题的立场。
结果，聊天机器人经常无法正确判断地址所属的国会选区，有时干脆拒绝回答。
上述引用比例来自这项特定问题和测试环境，并不意味着ChatGPT或Copilot在所有选举问题中都会以相同比例引用Ballot Brief。
《华盛顿邮报》发现，在9月的首轮测试中，Ballot Brief是ChatGPT回答此类问题时最常引用的来源。该报上周将结果告知OpenAI后，在第二轮测试中，ChatGPT引用该网站的比例降至4%。
OpenAI表示，公司希望通过可靠、权威的信息来源以及防范偏见和滥用的措施，帮助人们参与选举；当问题被指出时，公司会采取行动改善模型回答。《华盛顿邮报》在报道中披露，该报与OpenAI有内容合作关系。
Ballot Brief也是Copilot在测试中最常引用的来源。报道发表后，微软表示，已调整工具，让它优先使用权威来源和各州官方选举网站，并将继续测试和改进选举相关回答。
Anthropic的Claude在首轮测试中主要引用非营利组织运营的成熟选民指南，但在第二轮测试中开始频繁链接Ballot Brief。Anthropic没有解释这一变化，只表示公司已在中期选举前测试模型、加强防护措施，并努力提供可靠、准确的投票信息。
当AI开始引用AI写的内容
联网搜索让聊天机器人能够获取比训练数据更新的信息，也让它们更容易接触到互联网上由AI生成的内容。
美国西北大学博士研究人员穆瓦法克·阿拉哈姆（Mowafak Allaham）在一项研究中发现，针对政治、健康和环境等领域的问题，聊天机器人引用的来源中有**16%**本身由AI生成。
他表示，这些系统如何寻找、筛选和整合信息，并不透明。
另有两名曾负责卡玛拉·哈里斯2024年总统竞选数据分析的研究人员发现，聊天机器人在回答中期选举问题时，也高度依赖AI生成的网站。他们的测试还发现，长期提供选举信息的网站Ballotpedia此前一直在屏蔽ChatGPT。
Ballotpedia首席执行官莱斯利·格雷夫斯（Leslie Graves）表示，屏蔽是安全设置出错造成的，网站已于上周解除限制。Ballotpedia拥有约90人的团队，埃兹里克等人则希望借助AI，以更小的规模建立庞大的候选人数据库。
除Ballot Brief外，《华盛顿邮报》的测试还发现，聊天机器人经常引用DecodeTheVote和Smarter.Vote。前者使用AI角色介绍候选人的履历，后者还尝试预测选举结果。
Smarter.Vote创始人雅各布·卢科塔（Jacob Loukota）称，他的网站已被AI机器人引用170万次，远远超过人类访客的访问次数。他认为，聊天机器人喜欢引用该网站，可能是因为网站会标明信息来源。
卢科塔说，这并非他当初的计划，但他觉得这样的结果可以理解，网站的形式很适合AI聊天机器人使用。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
When United States voters asked ChatGPT or Microsoft Copilot for information about candidates, they might have been given answers from a small website they had never heard of. The site, known as Ballot Brief, collected information on thousands of candidates. Ben Torres Ezrick, its founder, wanted to help voters get better access to campaign information, as well as to attract a special type of visitor — the Al chat robot.
According to a Washington Post test, ChatGPT cited Ballot Brief in 22 per cent of the election-related responses, and Microsoft Copilot accounted for 75 per cent. There were several errors in the candidate information on this website, one of which was repeated to the questioner by Copilot.
With more and more Americans asking AI for political and electoral information, the survey raised a problem: When chat robots rely on websites created by AI, how do mistakes reach voters?
A small site, a common source of AI.
Ezrik held senior positions in Google, Zoom and the video platform Rumble. He found that voters often needed to spend a lot of time searching between different websites when they wanted to know about candidates, and decided to create a all-in-one website that provided election information.
Faced with information on thousands of candidates, he chose AI to help. He sent a number of numbers AI agent to gather information online and then put it in Ballot Brief for consultation.
The site attracts more visitors than humans. Ezrik states that the data available to him show that during the 30 days that passed between September and October, the Microsoft AI system cited Ballot Brief more than 680,000 times.
The Washington Post investigation found that Ballot Brief was only one of them. Three new websites, including it, often appear in voting sources provided by chat robots. The founders of the three websites indicated that their teams were small and almost entirely dependent on AI for data collection and summary.
The problem followed. According to Isabel Linzer, an electoral analyst at the Centre for Democracy and Technology, the results of the survey indicate that a person, publish, who generates information on the AI website, may be easily adopted by chat robots. She said that this could affect the confidence of voters in election information.
The error was repeated by AI.
The Washington Post found several errors on Ballot Brief.
One of the materials on the website stated that Jon Husted, a Republican senator in Ohio, had supported a bill. But that bill was passed before he took office. In reply to a question from the Washington Post, Copilot repeated this error.
The website also incorporated two California candidates, Robert Garcia, into the same document, and several other members of the House of Representatives were wrongly listed on the committee.
Some members of Parliament had rather brief information, lacked voting records, and the website had incorrectly indicated that they did not appear to have a campaign website.
Ezrik stated that he worked to ensure that information was accurate. After presenting his findings to the Washington Post, he made several adjustments to the way the website works.
The website, he said, was intended to provide voters with as much information as possible, but should not be the only source of information for them.
One test covering 3,100 districts
To understand how chat robots answer election questions, the Washington Post, in collaboration with the data company Promptwatch, tested the AI products of Open AI, Anthropic, Google, Microsoft, Amazon and Perpexity.
For about 3,100 counties in the United States, each county selected an address to ask the same question to chat robots and asked it to produce a November 2026 voter guide based on that address, listing all elections and candidates ' positions on key issues.
As a result, chat robots are often unable to correctly judge the parliamentary constituency to which the address belongs and sometimes simply refuse to answer.
The above-mentioned reference rates are derived from this particular problem and test environment and do not mean that ChatGPT or Copilot will be quoted in the same proportion in all election issues as Ballot Brief.
The Washington Post found that during the first round of tests in September, Ballot Brief was the most frequently cited source of questions from ChatGPT. After informing Open AI of the results last week, in the second round of tests, ChatGPT cited 4 per cent of the site.
Open AI stated that the company wanted to help people participate in the elections through reliable and authoritative sources of information and measures to prevent prejudice and abuse; when the question was raised, it would take action to improve the model answers. The Washington Post disclosed in its coverage that it had a content-related relationship with Open AI.
Ballot Brief is also the most frequently cited source of the test by Copilot. Following the publication of the report, Microsoft indicated that the tools had been adapted to give priority to authoritative sources and state official electoral websites and that it would continue to test and improve electoral-related responses.
Claude of Anthropic used the mature voter guide for non-profit organization operations in the first round of tests, but began to link Ballot Brief frequently in the second. Anthropic did not explain this change, but merely indicated that the company had tested models before the mid-term elections, strengthened protective measures and worked to provide reliable and accurate voting information.
When AI started quoting what it wrote
Network search allows chat robots to access information that is more up-to-date than training data, and also makes it easier for them to gain access to AI-generated content on the Internet.
In a study by Mowafak Allaham, a doctoral researcher at Northwestern University of the United States, it was found that **16 per cent of the sources cited by chat robots in relation to issues such as politics, health and the environment** were generated by AI itself.
He stated that there was no transparency in how those systems sought, filtered and consolidated information.
Two other researchers, who had been responsible for the analysis of Camara Harris ' 2024 presidential campaign data, found that chat robots were also highly dependent on AI-generated websites in answering mid-term election questions. Their tests also found that Ballotpedia, a long-standing website providing election information, had previously been blocking ChatGPT.
The Chief Executive Officer of Ballotpedia, Leslie Graves, stated that the shield was the result of an error in security and that the website had been released last week. Ballotpedia has a team of about 90 people, while Ezrik and others wish to use AI to build a larger pool of candidates on a smaller scale.
In addition to Ballot Brief, tests in the Washington Post found that chat robots often cited Deode TheVote and Smarter.Vote. The former uses the AI role to introduce candidates ' curricula vitae, while the latter also attempts to predict the results of elections.
According to the founder of Smarter.Vote, Jacob Loukota, his website has been quoted 1.7 million times by AI robots, far exceeding the number of visits by human visitors. In his view, chat robots preferred to quote the site, probably because it would indicate the source of information.
Lukota said that this was not his original plan, but he thought that the result was understandable, and that the site was well suited for the AI chat robot.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.