The Ad

On February 6, 2023, Google released an advertisement on Twitter for Bard, its conversational AI chatbot. Bard was Google's response to ChatGPT, which had launched in November 2022 and was generating enormous public and investor interest. Microsoft had moved to integrate ChatGPT's underlying technology into Bing, and Google was scrambling to show it could compete.

The ad described Bard as "a launchpad for curiosity" and showed the chatbot in action, responding to the prompt: "What new discoveries from the James Webb Space Telescope can I tell my 9-year old about?" Bard's response included several claims about JWST's discoveries: it took the very first pictures of a planet outside our own solar system. This was presented as one of several items a parent might share with a curious child.

The claim was factually wrong. The first direct image of an exoplanet was captured in 2004 by the European Southern Observatory's Very Large Telescope. The planet was 2M1207b, orbiting a brown dwarf roughly 170 light-years from Earth. The JWST, which began science operations in mid-2022, had made significant exoplanet discoveries, including the first direct image of an exoplanet captured by that particular telescope - but it was not the first telescope to do so.

The distinction matters because Google had chosen this prompt specifically to showcase Bard's ability to "simplify complex topics." And Bard simplified it by getting it wrong.

Reuters Breaks the Story

Reuters was first to identify the error, publishing its report on February 8, 2023 - the same day Google held a press event in Paris to demonstrate its AI plans. The timing was exceptionally bad. Google had organized the Paris event to show how AI would transform its search product, and the news cycle that day was supposed to be about Google's AI ambitions. Instead, the headline was that Google's AI chatbot had produced an inaccurate answer in its own promotional material.

This wasn't an edge case found by a researcher probing Bard's limits. This was the marketing team's hand-picked example of what Bard could do, published to millions of Twitter users as a demonstration of the technology's value. If the best example - the one Google selected, reviewed, and chose to publish - contained a factual error, the natural inference was that the typical response quality was worse.

$100 Billion in a Day

Alphabet shares fell roughly 9% on February 8, erasing approximately $100 billion in market value in a single trading session. The stock drop couldn't be attributed solely to the Bard error - investors were also disappointed by the Paris event, which was criticized as vague and unimpressive compared to Microsoft's concurrent announcements about integrating AI into Bing and Office.

But the Bard error was the lead story. NPR, CNN, the Guardian, Reuters, and Business Insider all reported on the factual mistake, and the stock drop was framed in nearly every outlet as a direct consequence. Gil Luria, a senior software analyst at D.A. Davidson, told Reuters: "While Google has been a leader in AI innovation over the last several years, they seemed to have fallen asleep on implementing this technology into their search product."

The market's reaction reflected a broader anxiety about Google's competitive position. ChatGPT's success had positioned Microsoft, through its partnership with OpenAI, as the frontrunner in consumer AI. Google had the underlying technology - its Transformer architecture was the "T" in GPT - but had been slower to bring a consumer product to market. When the product it did launch tripped on a basic factual question in its own ad, investors read it as confirmation that Google was behind.

The Anatomy of the Error

Bard's mistake was a textbook example of how large language models produce hallucinations. The JWST had indeed captured direct images of exoplanets - that was real. The model's error was in attributing the "first" to JWST rather than to an earlier telescope. The statement was close to true, structured like a true fact, and plausible enough that a non-expert might not catch it. But it was wrong.

This type of error - confident, specific, and subtly inaccurate - is exactly what makes AI hallucinations difficult to detect in practice. The answer wasn't gibberish. It wasn't a bizarre non sequitur. It was a factual claim about astronomy that required specialized knowledge to recognize as false. An astronomer or a space enthusiast would catch it immediately; a parent looking for things to tell their kid about space probably wouldn't.

The irony was that Bard's error was easy to fact-check. A Google search would have turned up the correct information in seconds. But Bard wasn't designed to do a Google search; it was designed to generate text that sounded like the answer to the question. And it did - the answer sounded authoritative, was well-structured, and was wrong.

The Competitive Context

February 2023 was the peak of the initial AI arms race between Google and Microsoft. Just one day before Google's Paris event, Microsoft had held its own event unveiling the new AI-powered Bing with ChatGPT integration. Microsoft CEO Satya Nadella was quoted saying he wanted to make Google "dance," and Google appeared to be doing exactly that - dancing, fast, and stepping on its own feet.

Google CEO Sundar Pichai had announced Bard in a blog post on February 6, emphasizing that the chatbot was powered by a "lightweight" version of Google's LaMDA model. The Paris event on February 8 was supposed to build on that momentum. Instead, the combination of the factual error in the ad and a reportedly underwhelming live demonstration reinforced the narrative that Google was rushing a product to market in response to competitive pressure.

Bard went on to launch publicly in March 2023 and was eventually renamed Gemini. By the end of 2023, Google had stabilized its AI product strategy and recovered its stock price. But the JWST moment remained a reference point - the time Google's answer to ChatGPT got a basic fact wrong on national television, in the company's own ad, on the day of its big event, and cost $100 billion before lunch.