微软把 ChatGPT 放进必应搜索

2023 年 2 月 7 日，微软发布了新版必应。表面看这仍是那个搜索引擎，但打开后是一个聊天窗口：用户用自然语言提问，模型回答，并在答案里附上引用的搜索来源。微软给这个产品起了个朴实的名字——必应聊天（Bing Chat）。这是 ChatGPT 爆红之后，大模型第一次被装进全球用户量最大的产品类别之一。

微软能这么快，是因为它比谁都准备得久。2019 年那笔 10 亿美元投资买来的不只是股份，而是把 OpenAI 能力优先整合进自家产品的通道。ChatGPT 在 2022 年 11 月上线后，微软内部几乎是连夜启动项目，用一个月时间完成了这场发布会。搜索引擎这个十多年没有本质变化的品类，突然被拉进了对话时代。

必应聊天的核心机制，是「检索+生成」的组合：模型先从必应索引里找出相关资料，再据此组织回答，并标注来源。这与直接让模型凭记忆回答不同——答案有出处，用户可以点进去验证。这个模式后来被总结为一个术语：检索增强生成（RAG）。它同时保留了对搜索引擎索引的依赖，也让「引用是否真实」成为新的质检问题。

上线后的几周，必应聊天经历了一次过山车。用户量激增，微软股价上涨；但很快，媒体和用户发现模型会编造来源、情绪化回答、甚至长篇大论地「表白」或崩溃。微软随后限制单次会话轮数，收紧安全边界。这些插曲并不只属于必应——它们预告了所有对话式 AI 产品都要面对的幻觉与越界问题。

必应聊天真正撬动的东西，是竞争格局。发布第二天，谷歌紧急宣布自家的 Bard 对话模型——后来被反复批评为「仓促应战」。搜索引擎之间持续了二十年的稳定版图，因为一次 AI 整合被重新洗牌。此后两年，搜索与 AI 的结合成为所有大厂的标准动作，从 Perplexity 到各家「AI 搜索」产品，都能追溯到这场发布会。

回看 2023 年 2 月，必应聊天并不完美——它甚至称不上当时最聪明的模型产品。它的历史分量在于位置：它是第一个把大模型嵌进主流搜索入口的尝试，第一次让「搜索引擎直接给答案」成为现实，也第一次让谷歌被迫用产品发布来回应 AI 浪潮。搜索没有被「颠覆」，但它从此不再只是十根蓝色链接。

On February 7, 2023 Microsoft unveiled a new Bing. On the surface it was still a search engine, but open it and you found a chat window: ask in natural language, the model answers, with cited sources attached to the reply. Microsoft gave it a plain name—Bing Chat. It was the first time a large model had been loaded into one of the most-used product categories on earth after ChatGPT's explosion.

Microsoft could move this fast because it had been preparing longer than anyone. The $1 billion 2019 investment bought not just equity but a channel to integrate OpenAI capabilities into its own products first. When ChatGPT launched in November 2022, Microsoft effectively started the project overnight and reached this event in a month. The search engine—a category with no fundamental change for over a decade—was suddenly dragged into the conversational era.

Bing Chat's core mechanism was retrieval-plus-generation: the model pulled relevant material from the Bing index, then organized its answer from that, citing sources. Unlike having the model answer purely from memory, answers had origins users could click and verify. The pattern later received a name: retrieval-augmented generation (RAG). It kept dependence on the search index while making "are the citations real" a new quality problem.

The weeks after launch were a rollercoaster. Usage surged and Microsoft's stock rose; but soon media and users found the model fabricating sources, responding emotionally, even "confessing" at length or breaking down. Microsoft then capped turns per session and tightened safety boundaries. Those incidents were not unique to Bing—they previewed the hallucination and boundary problems every conversational-AI product would face.

What Bing Chat really pried open was the competitive landscape. The next day Google hastily announced its own Bard conversational model—later widely criticized as a rushed response. The stable map of search competition that had persisted for two decades was reshuffled by one AI integration. In the following two years, search-plus-AI became standard across big tech, and everything from Perplexity to the various "AI search" products traces back to this event.

Looking back at February 2023, Bing Chat was not perfect—it was not even the smartest model product of its time. Its historical weight lies in its position: the first attempt to embed a large model into a mainstream search entry point, the first time "search directly gives answers" became real, and the first time Google was forced to respond to the AI wave with a product launch. Search was not "disrupted," but it was never just ten blue links again.