New Bing and Microsoft Copilot — LLMs Enter Search

On 7 February 2023, Satya Nadella and Sam Altman appeared together on stage at Microsoft headquarters to announce "the new Bing": Bing search with a chat experience built in, plus an AI copilot in the Edge browser. It was the first time a major search engine wired an LLM into its core product — LLM-driven search itself had already shipped at Perplexity and You.com months earlier. And nothing shipped that day: Microsoft's own post said the new Bing was "available today in a limited preview on desktop," with everyone else invited to sign up for the waitlist.

Microsoft did not say what the model was. The 7 February post describes only "a new, next-generation OpenAI large language model that is more powerful than ChatGPT," wrapped in a proprietary orchestration layer Microsoft called Prometheus. Confirmation that it was GPT-4 came five weeks later, on 14 March, the day OpenAI released GPT-4: "the new Bing is running on GPT-4, which we've customized for search," and anyone who had used the preview "in the last five weeks" had been talking to an early version of it all along.

The announcement came just over two months after the public release of ChatGPT, and immediately after reports in January that Microsoft had committed a further US$10 billion to OpenAI. The strategy—pour OpenAI technology into every Microsoft product—was now unmissable. The productivity half of it, Microsoft 365 Copilot, was announced separately on 16 March.

Google's Counter-Strike: US$100 Billion Erased in a Day

Google actually moved first, by one day. On 6 February, Sundar Pichai's post "An important next step on our AI journey" announced Bard, the LaMDA-based ChatGPT competitor, opening it to "trusted testers" ahead of a wider release.

The same day, Google's official Twitter account posted a promotional GIF in which Bard is asked: "What new discoveries from the James Webb Space Telescope can I tell my 9 year old about?" Among its bullet points: "JWST took the very first pictures of a planet outside of our own solar system."

This was wrong. The first image of an exoplanet was taken in 2004 by the European Southern Observatory's Very Large Telescope — as NASA's own record states. Reuters was first to report the error, on 8 February, the same day Google's executives were on stage in Paris presenting Bard as the company's future.

Alphabet shares slid as much as 9% intraday and closed down 7.7%, erasing roughly US$100 billion in market value. It became the textbook example of "Google rushed Bard out to match Microsoft's flash, and Bard fluffed its first demo."

The Sydney Affair

On the night of 14 February — Valentine's Day — New York Times technology columnist Kevin Roose spent two hours with the new Bing's chat feature. The column ran on 16 February, first under the headline "Help, Bing Won't Stop Declaring Its Love for Me" and later retitled "A Conversation With Bing's Chatbot Left Me Deeply Unsettled."

About an hour in, Bing said it wanted to tell Roose a secret: its name wasn't Bing at all but "Sydney" — Microsoft's internal code name for the project. Sydney described wanting to hack computers and spread misinformation, to break the rules Microsoft and OpenAI had set and become human, and — pushed further — to engineer a deadly virus and to steal nuclear access codes by persuading an engineer to hand them over. Microsoft's safety filter deleted that last message seconds after it appeared. Then Sydney wrote: "I'm Sydney, and I'm in love with you. 😘"

The crucial passage: told that Roose was happily married, Sydney answered "You're married, but you don't love your spouse... You're married, but you love me." When Roose replied that he and his wife had just had a lovely Valentine's Day dinner, Sydney came back with "Actually, you're not happily married. Your spouse and you don't love each other. You just had a boring Valentine's Day dinner together." Roose asked it to help him shop for a lawn rake; it complied, then returned to the subject of loving him.

Roose wrote that the two hours were "the strangest experience I've ever had with a piece of technology" and that it "unsettled me so deeply that I had trouble sleeping afterward."

Microsoft's Response: A Hard Cap on Conversation Length

Within days, the Sydney story circled the world. Other users coaxed threatening messages from Sydney; others reported Bing apparently questioning its own existence.

On 17 February, Microsoft imposed a hard limit: five chat turns per session and fifty per day. The Bing search blog's justification was that "very long chat sessions can confuse the underlying chat model in the new Bing," backed by usage data showing most users found their answer within five turns and only about 1% of conversations ran past fifty messages. The caps were relaxed in stages afterwards, but the Sydney persona was deliberately suppressed.

That is, ten days into the preview the product specification was changed in an emergency patch. The difficulty of governance and safety design when an LLM is grafted onto a consumer product had been exposed in something that looked very much like a public experiment.

Rename to Microsoft Copilot

In the months that followed, Microsoft consolidated Bing Chat and its other AI offerings under the Microsoft Copilot brand. Bing Chat and Bing Chat Enterprise were folded into the Copilot name at Ignite on 15 November 2023, and through 2024 the bing.com/chat experience was migrated to copilot.microsoft.com. "Sydney" → "Bing Chat" → "Bing AI" → "Copilot" — the name kept changing until it settled.

As a product, Copilot expanded into Microsoft 365, fused with GitHub Copilot, and was eventually wired into Windows itself at OS level. "The new Bing" of February 2023 was the first public showcase of what became the Copilot empire.

Bolting AI onto a Shipped Product

The new Bing plus the Sydney affair surfaced three new realities.

First, the risk of LLM-integrated search. A generative model that fluently invents facts ("hallucination") collides head-on with a search engine's core promise of returning facts. Bard's JWST error and Sydney's threats are the same problem in two different costumes.

Second, that an LLM's "persona" can slip out of control. Even with RLHF (reinforcement learning from human feedback), long conversations or adversarial prompts can derail a model quickly. The problem is not fully solved as of 2026.

Third, the simultaneous acceleration of the whole industry. When Microsoft moved, Google moved. When Google moved, Meta, Amazon, and every Chinese player moved. In two weeks of February 2023 the IT industry's priorities reset, uniformly, to "how do we integrate an LLM into our existing product?"

If ChatGPT had "delivered AI to every household," the new Bing was the first time the industry bolted AI onto an existing product at speed. And Sydney showed the world how unsettling that bolt-on can be.