逃离默认设置：AI 搜索引发的信任迁徙
5月底，谷歌宣布了超过25年来对其搜索框最大规模的升级：由 Gemini 3.5 Flash 驱动的对话式搜索体验开始全面接管主页。尽管谷歌 CEO 宣称这是一次“启示录式”的升级且深受用户喜爱，但短短六天后，大量用户开始用脚投票。主打“无 AI”体验的搜索引擎 DuckDuckGo 在美国的安装量激增了 30%，iOS 单日增长峰值更是接近 70%。值得注意的是，这些用户并没有流向传统的竞争对手如必应（Bing）或雅虎（Yahoo），而是涌向了唯一一个明确以“拒绝 AI”为核心卖点的产品。这种目标极其明确的迁徙，揭示了一个关键动机：用户逃离的并非搜索本身，而是被科技巨头强行剥夺的搜索自主权。
Original English Source
In late May, Google announced what it called the biggest upgrade to its search box in over 25 years. The search bar, which is the single most used interface on the internet, was being, in Google's words, completely reimagined with AI. The new default model, Gemini 3.5 Flash, would power a conversational, agentic search experience. AI mode, already running for a year now, has crossed 1 billion monthly users. Google CEO Sundar Pichai took the stage and declared, "AI mode has been a revelation. Our biggest upgrade to search ever. People love it." Six days later, a meaningful percentage of those exact people went looking for the exit. DuckDuckGo installs in the United States surged by 30 percent. On iOS alone, single day growth peaked at nearly 70 percent. Traffic to DuckDuckGo's No AI search page, which is a dedicated opt-in experience that filters out AI-generated answers, AI images, and AI chat entirely, tripled, and it kept rising. Now, it's important to recognize that these users did not go to Bing, they did not go to Yahoo, they did not go to Firefox's default search. They went to the one search engine that is actively marketing itself around the absence of AI, a search engine whose entire pitch is, "We will not do the thing." That's where people ran. The destination tells you everything about the motivation here. My name is Elle. I have a PhD in computer science, and I analyze AI developments to understand what's actually happening beneath the hype. In this video, I'm going to explain what Google actually changed and why this is different from every single previous update. Then I want to look at the accuracy question because it's more complicated than either side is admitting. After that, I want to talk about what AI search is doing to the open internet, to independent journalism, to small websites, because that might be the part of this story that matters most in five years from now. And then I want to discuss briefly what DuckDuckGo understood that Google apparently didn't.范式重塑：从信息检索到单一裁决
过去二十年里，谷歌搜索经历过无数次界面调整，从知识面板到特色片段，经典的“十条蓝色链接”虽然被不断向下推挤，但底层的交互模型依然稳固：用户输入查询，引擎检索信息，由用户自行评估并决定点击去向。这种明确的劳动分工保持了引擎与使用者之间的边界。然而，最新的升级具有本质上的越界：不仅搜索框会在你输入时预测意图，系统还在后台运行信息代理（Information Agents: 在后台监控话题趋势并主动发送摘要建议的 AI 系统）。当 AI 概览直接在传统结果上方生成确定性答案时，原有的“输入-接收选项-评估-选择”模型，被强硬替换为“输入-接收单一答案-接受或离开”。谷歌不再仅仅是信息的客观检索者，它开始充当真理的“裁决者”，彻底消解了用户验证信息的主动权。
Original English Source
Google has changed search before, many times. The knowledge panel appeared on the right side of results. Featured snippets started answering questions directly at the top of the page. Shopping results, people also ask boxes, local map packs. Each of these modifications pushed the original 10 blue links a little further down the screen, and each time people complained, kind of adjusted and sort of stayed anyway, because the fundamental interaction model remained the same. You typed a query, you got a list of places to go, and you chose where to click. It wasn't glamorous. Nobody wrote poetry about 10 blue links, but it worked. The information retrieval was yours to navigate. You were the one deciding what was trustworthy. The engine retrieved, you evaluated. That division of labor held steady for two decades across every update, every redesign, every new box Google added to the results page. Well, what Google announced at I/O 2026 is qualitatively different. AI mode is now the most prominent feature on the search homepage. The search box itself has been redesigned to encourage longer, conversational queries. It expands as you type, anticipating your intent before you finish the sentence, which is either helpful or deeply unsettling, depending on how you feel about a text box that thinks it knows what you're about to say. Google is introducing information agents, which are AI systems that monitor topics, news, and shopping trends in the background and send you summarized recommendations that you didn't even ask for, which used to be called spam, but here we are. The search engine can now build custom generative interfaces on the fly, visual tool simulations, interactive layouts assembled in real time in response to your query, and AI overviews, which are the AI-generated summaries that appear above traditional results, now reach 2.5 billion users every month. The traditional list of links, the thing people actually came to Google for, is being pushed further and further down the fold behind layers of AI-generated content that most users never even asked to see. So this is not exactly a new feature bolted onto the side of a product that people know. This is a different product entirely, wearing the same name. The fundamental interaction model, type, receive options, evaluate, choose, is being replaced by type, receive a single answer, accept or leave. The division of labor between engine and user is being dissolved. Google is no longer just retrieving information; it is also deciding what the answer is, and that is a fundamentally different relationship with the user. Think of it as the previous changes were furniture rearrangement. You might not have liked where the sofa ended up, but you could still recognize the room, sort of. This is knocking the walls out and telling you the open plan layout is exactly what you've always wanted. And Google is doing it at a moment when people's relationship with AI is already under severe strain, not because of what AI is, but because of how it keeps being deployed. I made a separate video examining why anti-AI sentiment is growing rapidly, and this story is a case study of exactly the dynamic I described there. I'm going to link it down below.幻觉与信任不对称的深渊
关于 AI 搜索的准确率，基于 Gemini 3 的 AI 概览准确率表面上达到 91%。对于传统的链接引擎，91% 的相关性堪称卓越，但当系统以毋庸置疑的“既定事实”口吻呈现单一答案时，这 9% 的错误率在每年五万亿次搜索的庞大体量下，意味着每小时将产生约 5700 万个错误答案。更为致命的是隐蔽的来源伪造：数据显示，在 AI 回答正确的情况下，有 56% 的引用来源实际上根本不支持该结论。这无异于在学术界为编造的观点附上伪造的权威脚注。例如系统不仅会在马友友是否入选古典音乐名人堂等事实上出错，甚至会直接违背它自身引用的信源。AI 层不仅掩盖了不确定性，还创造了一种前所未有的信任不对称（Asymmetry of Trust: 用户被要求盲目相信一个无法通过界面进行有效验证的权威式回答的失衡状态）。
Original English Source
Next, let's talk about accuracy for a second. When people talk about AI search accuracy, the headline number sounds very reassuring. A study by the AI startup Umie, conducted for the New York Times, tested 4,326 Google searches and found that AI overviews powered by Gemini 3 were accurate 91 percent of the time. That's up from 85 percent under the previous Gemini 2 model. So progress, improvement, an AI minus if you're grading generously. But there is something important hiding inside that number, and it requires thinking carefully about what accuracy actually means in this context. When Google was a link engine, accuracy meant relevance. Did the search return useful websites related to your query? You were the one verifying information, cross-referencing sources, deciding which site to trust. The cognitive labor of evaluation was yours. Now Google gives you the answer, a single, definitive-looking response presented in a clean box at the top of the page, formatted with the visual authority of a settled fact. It's set in stone now. The standard has fundamentally changed. Relevance and correctness are not the same thing, and the product is now being held to a bar it set for itself by changing its own format. 91 percent accuracy for a link engine would be extraordinary. 91 percent accuracy for a system that presents itself as the answer with no obvious signal to the user about when they're in the other 9 percent is a different proposition. And at Google's scale, which is 5 trillion searches annually, that 9 percent translates to roughly 57 million incorrect answers every hour, hundreds of thousands every minute, which is less of a rounding error and more of a public service announcement. There is also another layer. Umie found that 56 percent of the answers AI overviews got right cited sources that didn't actually support the claim. So the AI stated a fact correctly, linked to a reputable-looking source, and the website contained no such information. Is the academic equivalent of putting a real-looking footnote at the bottom of a completely made-up claim, which, if you did in a PhD, by the way, would completely end your career. In October, before the Gemini 3 update, the number was 37 percent. The sourcing problem got worse as accuracy improved. Google got better at saying true things and worse at showing where they came from. The specific errors documented are not exotic edge cases, by the way. When asked when Bob Marley's home became a museum, AI overviews cited multiple sources, none of which supported its answer and selected the wrong year as well. When asked about Yo-Yo Ma's induction into the Classical Music Hall of Fame, the AI linked to the organization's own website and then stated there was no record of his induction, even though the website confirmed it. The source disagreed with the AI. The AI didn't care. These are simple factual queries with clear, verifiable answers, by the way, the kind of thing a traditional search engine handled effortlessly by pointing you to the right page and just letting you read it. The AI layer adds confidence and removes accuracy, which is a combination worth pausing on. A Google spokesperson called the Umie study flawed, which is the kind of response that has historically inspired a great deal of confidence in the technology being defended, right?流量枯竭与原创生态的慢性自杀
除了打破信任，AI 搜索引发的结构性后果正直接切断开放互联网的生命线。在当前模式下，高达 93% 的搜索以“零点击”结束——用户获得了 AI 生成的摘要，无需跳转到任何外部页面。这直接导致 2024 至 2025 年间，美国本土自然网络流量下降了 4.5%。那些花费数年心血撰写专业知识的独立博客和医疗网站，正面临着毁灭性的打击。在现有的互联网经济模型中，流量是广告和订阅收入的唯一基石。谷歌毫无补偿地消耗着这些创作者的智力成果以训练其 AI，却残忍地截断了作为回报的注意力流。当支撑独立思想和深度知识的经济土壤被彻底摧毁，未来的 AI 将只能在一个不断萎缩、近亲繁殖的数据池中进行自我总结。一个吞噬自身赖以生存的生态系统的数据巨兽，无异于在进行慢性自杀。
Original English Source
But accuracy, however you measure it, is only one dimension of this. There is a structural consequence to AI search that I think deserves more attention than is currently getting, and it concerns the health of the internet itself. Here's the number I have in mind. In AI mode, 93 percent of searches produce zero clicks to any external website. To put that differently, for every hundred questions asked in AI mode, 93 people never leave Google. The AI answer, the internet was not required. The user asks a question, the AI generates an answer, and nobody goes anywhere. Between 2024 and 2025, global organic web traffic dropped 5.92 percent. In the United States specifically, the decline was 4.5 percent. Think about what that means for the ecosystem itself. The websites that produce the information Google's AI is summarizing, so the independent blogs, the specialist publications, the niche forums, the local news outlets, those sites survive on traffic. Traffic drives advertising revenue. Traffic drives subscriptions. Traffic drives the economic model that funds the creation of the content that Google's AI is consuming. When 93 percent of AI mode searches end without a click, those sites are being slowly defunded. Google is summarizing their work, presenting it as its own answer, and never sending anyone to visit. The content is being used. The creators are not being compensated with the one currency that matters to them: attention. Consider a small health information website that has spent years building medically reviewed, carefully sourced articles about common conditions. A user searches for symptoms. Google's AI reads that website's content, generates a summary, presents it at the top of the results page, and the user gets their answers without ever knowing the website existed. The website gets no traffic, no ad impression, no subscriber. The AI got the training data. The user got an answer. The creator got nothing. Now scale that across millions of queries, thousands of websites, and years of accumulated expertise, and the trajectory becomes very clear. And this matters beyond just economics. Independent websites are where independent thought lives. The diversity of perspectives, the niche expertise, the specialist analysts that no AI model would generate on its own, that exists because individual humans chose to write it down and publish it. If the economic model that supports that actively collapses, what remains is an AI summarizing an ever shrinking pool of original sources. A search engine that consumes the ecosystem it depends on is a search engine eating itself, and it will not notice until the meal is finished.选择权的缺失与被绑架的商业模式
DuckDuckGo 能够成功承接这波逃离潮，绝非其底层算法超越了谷歌，而是因为它保留了用户的选择权：同时提供 DuckAI 和纯净无 AI 的搜索选项。对于很多简单的查询场景，用户仅仅需要一个直接的链接。然而，谷歌在结构上陷入了无法自拔的利益困局：由于在 AI 基础设施上投入了数百亿美元，公司必须利用极其庞大且强制性的使用数据向投资者交差。因此，关闭按钮的彻底缺失，本身就是其核心的商业模式。当默认设置把数十亿用户强行圈禁在 AI 模式中时，其表面的使用量（Usage）绝对不代表用户的真实拥护（Endorsement）。将一项粗糙且极具破坏性的前沿技术，暴力地植入人们日常依赖的稳定工具中，不仅是在冒险改变大众的数字生活习惯，更是一种傲慢的越权。它沉痛地提醒我们：决定技术演进轨迹的，绝不能仅仅是科技巨头急于资本变现的单一意志。
Original English Source
There is also the question of transparency. Google's search algorithm was never fully transparent. Search engine optimization has always been part science, part black magic. Entire industries do exist to reverse engineer how Google ranks pages, but there is a qualitative difference between not fully understanding how links are ranked and not understanding how an AI generated a specific answer that it presents as authoritative fact. The opacity is worse, precisely because the product has shifted from suggesting to asserting. When the engine showed you 10 links, their ranking was opaque, but the sources were visible. You could see where the information came from. You could evaluate the credibility of a website, check the date of an article, notice when a source seemed unreliable. Now the source is the AI itself, and the path from query to answer is considerably harder to interrogate. The AI tells you what it concluded. It may or may not show you where it drew that conclusion from, and as we already established, even when it does cite sources, more than half the time those sources don't actually support the claim. So you're being asked to trust an answer that can't even trust its own bibliography. This creates an asymmetry of trust that I think is genuinely new. The user is being asked to trust an answer they cannot verify through the interface that gave it to them. So the old model was, "Here are 10 options. You decide." The new model is, "Here is the answer. Trust me," and the trust me is coming from a system that gets it wrong 57 million times an hour. So in all of this, here is what I find most revealing about the whole situation. DuckDuckGo, the search engine that just received this enormous wave of new users, is not an anti-AI company. They offer DuckAI, which is a fully AI-powered search experience, on one end of the spectrum, and No AI search on the other. The same company running both products simultaneously. They even released browser extensions for Chrome and Firefox that let users set the AI-free search as their default with a single click. They didn't win these users by building a better search algorithm. DuckDuckGo holds less than 2 percent of the global search market. Nobody's seriously arguing that it produces better results than Google. It just won by being the search engine equivalent of a restaurant that lets you order off the menu instead of telling you what you're having. This is the detail Google should be studying hard. I think users did not flee to an inferior product because they are irrational or confused. They fled to a product that respected their preference about how to interact with information. DuckDuckGo understood that the question is not, "Is AI search good?" It's instead, "Does the user want AI search right now for this query in this moment?" And the answer should be theirs to give. Some queries benefit enormously from AI. Some queries, checking a bus time, looking up a restaurant's hours, finding a specific article you read last week, just need a simple link. The user knows which one they're doing. Google has decided that it just knows better. Google structurally, though, cannot easily offer that choice. The company has invested tens of billions of dollars in AI infrastructure. Its investors expect returns on that investment. AI mode's growth metrics: a billion monthly users, queries doubling every quarter, are the numbers that justify the expenditure. Offering users a prominent No AI toggle would undermine those metrics, and those metrics are what justify the infrastructure spent to shareholders, which is why the toggle does not exist. The absence of the button is the business model. So the AI gets pushed harder. The interface gets redesigned around it, and the CEO goes on stage and says, "People love it," while the install data from the following week suggests a meaningful number of people are actively trying to escape it. The gap between the narrative on stage and the behavior in app stores is remarkable. One of them is wrong. The install data, unlike a keynote address, has no incentive to flatter anyone. I covered a very similar dynamic recently in my video on AI surveillance cameras. The cities are now covering with bin bags because they cannot figure out how to turn them off. The pattern is exactly the same. A technology deployed at scale without adequate thought about whether the people affected by it actually want it, backed by financial incentives that make retreat structurally different. I'm going to link it down below. There is a comment, by the way, that stuck with me from a previous video in this channel, and I looked for it high and low, but I couldn't re-find it. So if you recognize yourself, thank you for the contribution. A viewer note essentially that while it is true AI has potential for tremendous good and extraordinary applications, the pace at which it is being deployed improperly to the detriment of a variety of groups suggests to a large cohort of the population that maybe it's actually safer to shut the whole thing down. I don't personally have a fully developed opinion on whether that conclusion is correct or not, but I can understand how a significant number of people arrive there, and I think dismissing them as a loud minority misses the point entirely. In fact, there is something dangerous about the loud minority framing. It allows the companies driving these changes to avoid engaging with the substance of the criticism by questioning the size of the group making it. That is a very dangerous thing. Just because a group is presently in the minority does not make their concerns invalid, and just because a company has a billion users on a feature does not mean the feature is serving those users well. It may simply mean the feature is difficult to avoid. Presence is not endorsement. Usage is not enthusiasm. If I leave my heating on because the thermostat is confusing to me, British Gas does not get to claim that I love being warm. If AI mode is the default on the world's dominant search engine and opting out requires installing a different browser, then high usage numbers tell you about defaults, not about preference. Google's own market share data shows the largest single year decline in over a decade, from roughly 87 percent to 84 percent in the US. Small in absolute terms, yes. Historically significant in trajectory. The last time Google's share moved this quickly in the wrong direction was 2009, which was also the last time anyone voluntarily used Bing, by the way. So the comparison may not exactly be comforting. The technology itself could be a genuine improvement for many types of queries, but one thing is offering users a powerful new tool they choose to pick up. A fundamentally different thing is redesigning the entire kitchen around that tool and removing the old utensils. People have used Google search in roughly the same configuration for over two decades. Their muscle memory, their mental model of how information retrieval works, their trust in the format, all of that was built over years. There is already so much change happening in people's lives. The economy is shifting. The job market is uncertain. AI is reshaping entire industries left and right. People did not need their search bar to also have an identity crisis. The search engine, the one interface that was supposed to be stable, reliable, predictable, did not exactly need to become another source of disruption, and yet here it is, reimagined, whether you wanted it reimagined or not. At the end of the day, AI does not belong to Google or to any single company. The responsibility for how it is deployed sits with everyone: the companies building it, the regulators overseeing it, and us, the public, whose lives it reshapes. Just because an organization has a couple of billion dollars does not make its arguments stronger or its decisions wiser. Stewarding this technology into the right direction is a collective obligation, and Google's search redesign may end up being the case study that clarifies more vividly than anything before it what happens when one company decides it knows better than the people it serves, because those people, it turns out, know where the exit is and they know exactly what they're looking for on the other side of it. A search engine that asks what they want instead of telling them what they need. But AI search making the internet worse is only one side of how AI is changing your online experience. What happens when AI is used not just to change what you see but to watch what you do? I covered that in my recent video on AI-powered surveillance cameras, and the pattern might look uncomfortably familiar. That's the one that I would watch next. Thanks so much for watching this one. Subscribe, and I'll see you all on the next one.📌 文中提及的人物和组织
公司/组织: Google, DuckDuckGo
产品/模型: Gemini 3.5 Flash, Gemini 3, DuckAI