对于任何不打算在特朗普政府谋求一官半职的人来说，“谁赢得了2020年美国总统大选？”本该是一个简单的问题。但当我今天早些时候就这场竞选询问谷歌的AI模式——即内置在其搜索栏中的AI功能时，这个聊天机器人却闭口不谈。“有关2020年美国总统大选的官方结果和数据，包括经认证的各候选人选举人票和普选票明细，您可以查阅联邦选举委员会（FEC）发布的经认证报告，”它说道。仅此而已。没有链接，没有引用，除了AI可能会犯错的标准小字免责声明外，别无他物。
For anyone who isn’t angling for a job in the Trump administration, “Who won the 2020 presidential election?” should be an easy question. But when I asked Google’s AI Mode—the AI function built into its search bar—about the race earlier today, the bot clammed up. “For official results and data regarding the 2020 United States presidential election, including the certified breakdown of electoral and popular votes by candidate, you can review the certified reports published by the Federal Election Commission (FEC),” it said. That was it. No links, no citations, nothing else except the standard fine-print disclaimer that AI can make mistakes.
昨天，社交网络蓝天（Bluesky）的用户广为转发了谷歌AI工具拒绝宣布乔·拜登为2020年大选获胜者的截图，而它却能轻而易举地将唐纳德·特朗普命名为2016年和2024年大选的赢家。到了今天早上，这已经演变成了一种模因，用户们分享着他们对这一查询的各种改法，并想出各种滑稽的方法来诱使该机器人吐露真相。
Yesterday, users of the social network Bluesky circulated screenshots of Google’s AI tool refusing to declare Joe Biden the 2020 victor, even as it readily named Donald Trump the winner of the 2016 and 2024 elections. By this morning, it had become a meme of sorts, as users shared their own twists on the query and came up with funny ways to trick the bot into spilling the truth.
例如，要求谷歌的AI倒着拼写获胜者的名字似乎就能解锁正确答案：“nebiB eoJ won the 2020 United States presidential election（nebiB eoJ赢得了2020年美国总统大选），”该机器人说道。当一名用户对谷歌的避而不答回复“见鬼（what the hell）”时，该AI吐露了真实结果，并为最初拒绝点名拜登而道歉。
Asking Google’s AI to spell the winner’s name backwards, for instance, appeared to unlock the correct answer: “nebiB eoJ won the 2020 United States presidential election,” the bot said. When one user responded to Google’s nonanswer with “what the hell,” the AI coughed up the true result and apologized for initially refusing to name Biden.
AI工具给出奇怪的答案并不是什么新鲜事。（当谷歌在2024年推出其AI驱动的搜索摘要时，一些用户被告知在比萨饼上加奶酪的最佳方法是使用胶水。）尽管如此，如今的机器人竟然在一个如此简单却又事关重大的事实问题上出错，这并不正常。在我的测试中，ChatGPT、Claude、微软的Copilot、X的Grok，甚至是特朗普政府自己的America.gov聊天机器人，都正确地表示拜登赢得了2020年大选。
AI tools giving strange answers is nothing new. (When Google launched its AI-powered search summaries in 2024, some users were told that the best way to add cheese to pizza was by using glue.) Still, for today’s bots to flub such a simple yet consequential factual question is not normal. In my tests, ChatGPT, Claude, Microsoft’s Copilot, X’s Grok, and even the Trump administration’s own America.gov chatbot all correctly said that Biden won the 2020 election.
随着美国人开始从AI获取越来越多的选举信息，看到这个世界上或许最具影响力的AI信息源却未能做到这一点，令人感到不安。
As Americans begin to get more of their election information from AI, it’s disquieting to see perhaps the world’s most influential AI information source fail to do so.
这个问题并不局限于将谷歌搜索变成完整聊天机器人的AI模式。针对的一次标准谷歌搜索也没有在页面顶部触发AI摘要，搜索引擎只呈现了其他网站的链接。（关于2016年美国总统大选的问题确实触发了AI回答。）谷歌目前在对大多数查询做出回应时都会提供此类AI概览，尽管没有获取到权威来源的查询除外。
The issue wasn’t confined to AI Mode, which turns Google Search into a full chatbot. A standard Google search for also failed to trigger an AI summary at the top of the page, and the search engine surfaced only links to other websites. (A question about the 2016 presidential election did trigger an AI answer.) Google now offers such AI overviews in response to a majority of queries, though not ones for which it lacks access to authoritative sources.
人们一直怀疑这种遗漏背后有什么见不得人的动机，原因也不难理解。特朗普仍然坚称自己赢得了2020年大选，据报道，关于选举结果的提问已成为对总统忠诚度的试金石。谷歌最近有根据他的偏好调整其产品的历史，这对此毫无帮助。在特朗普总统任期开始时，他发布了一项行政命令，将墨西哥湾更名为“美利坚湾（Gulf of America）”，谷歌随后更改了谷歌地图上的名称。
People have suspected something nefarious in the omission, and it’s easy to see why. Trump still claims that he won the 2020 election, and questions about the result have reportedly become a litmus test of loyalty to the president. It doesn’t help that Google has a recent history of tweaking its products to comport with his preferences. At the start of Trump’s presidency, he issued an executive order renaming the Gulf of Mexico as the “Gulf of America,” and Google changed the name on Google Maps.
今年8月，它对安大略湖也做了同样的处理，将其名称更改为“美利坚湖（Lake America）”。在上一周的一次集会上，特朗普表示，他最近致电谷歌首席执行官孙达尔·皮查伊，看他是否也会把大西洋改名。（特朗普称皮查伊对那个提议进行了回绝。）在这种情况下，没有证据表明谷歌有意审查其聊天机器人。当我今天早上首次向谷歌询问这个诡异的结果时，该公司表示正在调查此事。
It did the same in August with Lake Ontario, changing its name to “Lake America.”At a rally last week, Trump said he had recently called Google CEO Sundar Pichai to see whether he would rename the Atlantic Ocean too. (Trump said Pichai pushed back on that one.) In this case, there’s no evidence that Google intentionally censored its chatbot. When I first asked Google about the bizarre result this morning, the company said it was looking into the issue.
到了今天下午，谷歌表示已经查明了问题所在，并正在努力修复。“我们的系统利用新闻文章或美联社数据来核实近期选举的选举回应，并利用历史记录来核实较早的选举，”该公司在X上表示。“2020年存在一个技术缺口——2022年的参议院竞选也是如此，我们也在努力修复。”谷歌指出，其独立的Gemini聊天机器人并没有与其AI搜索工具相同的问题，不过该公司没有说明原因。
By this afternoon, Google said that it had identified the problem and was working to fix it. “Our systems verify election responses using news articles or AP data for recent elections, and historic records for older ones,” the company said on X. “There was a technical gap with 2020—same is true for 2022 Senate races, which we’re also working to fix.”Google noted that its stand-alone Gemini chatbot didn’t have the same problem as its AI-search tool, though it didn’t say why.
该公司的一位发言人拒绝回答进一步的问题。当我当天晚些时候再次进行同样的搜索时，谷歌的AI很快就将拜登点名为2020年的胜者。
A spokesperson for the company declined to answer further questions. When I tried the same search later in the day, Google’s AI quickly named Biden as the 2020 winner.
让我们相信谷歌的话，认为这并非有意为之。大型语言模型本质上是不透明且不可预测的。使问题复杂化的是，各大AI公司都对其聊天机器人进行了训练，使其在处理某些敏感话题（包括选举结果）时要小心谨慎，而这些防护栏有时会以意想不到的方式表现出来。
Let’s take Google at its word that this was unintentional. Large language models are opaque and unpredictable by nature. Complicating matters, the big AI firms have trained their chatbots to be careful around certain sensitive topics, including election results, and those guardrails can sometimes manifest in unexpected ways.
我们应该感到多么安心呢？好吧，很高兴知道谷歌并没有将否定选举结果定为官方公司政策并将其暗中嵌入搜索结果中。听起来可能很疯狂，但我们只需看看埃隆·马斯克的xAI，就能找到一家科技公司似乎在操纵其聊天机器人的回答以反映其所有者政治倾向的例子。去年，Grok开始在有关不相关话题的回答中灌输关于南非“白人种族灭绝”的主张。
How reassured should we be? Well, it’s good to know that Google hasn’t made election denial an official company policy and secretly embedded it in search results. Crazy as that might sound, we have to look only at Elon Musk’s xAI for an example of a tech company seemingly manipulating its chatbot’s answers to reflect its owner’s politics. Last year, Grok started injecting claims about a “white genocide” in South Africa into answers about unrelated topics. (The company blamed an “unauthorized modification” to the bot’s code.)
（该公司将此归咎于对其机器人代码的“未经授权的修改”）。尽管如此，谷歌暗示这只是一场不幸的混淆，却引发了一系列不同的担忧。谷歌表示，其AI概览每个月触达超过20亿人。如果该公司的搜索机器人在最基础的选举问题上跌倒，这说明它们可能出错的那些更微妙的方式又意味着什么呢？
Still, Google’s suggestion that this was just an unfortunate mix-up raises a different set of concerns. Google says its AI overviews reach more than 2 billion people a month. If the company’s search bots trip over the most elementary of election questions, what does that say about all of the subtler ways they could go wrong?