Google Bard Launch
One wrong sentence in a marketing GIF cost Alphabet $100 billion in a single day
What They Said
On February 6, 2023 — two days before a Paris event Google had hastily scheduled to respond to Microsoft’s Bing Chat announcement — the company published a blog post introducing Bard, its conversational AI. The post included a marketing GIF showing Bard answering the prompt: “What new discoveries from the James Webb Space Telescope can I tell my 9-year-old about?”
Bard’s answer named three discoveries. One of them claimed that JWST “took the very first pictures of a planet outside of our own solar system.” The post was timed to neutralize the narrative that Google had been caught flat-footed by OpenAI and Microsoft. CEO Sundar Pichai personally signed the announcement.
What Actually Happened
Astronomers noticed within hours. The first image of an exoplanet was captured in 2004 by the European Southern Observatory’s Very Large Telescope — 18 years before JWST began operations. The error was not subtle and not contested. Reuters confirmed it with NASA. By the morning of February 8, 2023, the day of the Paris keynote, the mistake was the lead in technology coverage worldwide.
The keynote itself compounded the damage. Google’s presenters declined to take live questions, did not demonstrate Bard answering novel prompts on stage, and provided no timeline for public access. Investors interpreted the staged event and the pre-canned demo containing a basic factual error as evidence that Google’s AI capabilities were not where the company had implied. By the close of trading on February 8, Alphabet stock had fallen 7.7%, with intraday losses approaching 9%, erasing roughly $100 billion in market capitalization in a single session.
The narrative damage outlasted the stock recovery. For most of 2023, Google was framed in the press as the AI incumbent caught off guard, even as Gemini was being prepared. The Bard brand never recovered and was retired in early 2024 in favor of Gemini. Internally, Pichai sent a company-wide memo asking employees to spend two-to-four hours testing Bard themselves — a tacit acknowledgment that the launch had skipped basic dogfooding.
The Root Cause
Google launched a marketing asset before anyone fact-checked the model’s output. The JWST claim was the kind of error a single astronomer or a single Google search would have caught. Nobody ran the check. The blog post and the GIF were treated as marketing collateral on a deadline rather than a public claim by the model itself. The deadline was set by Microsoft’s launch window, not by Bard’s readiness.
The second failure was the staged demo. A pre-recorded answer in a launch GIF carries the same accuracy expectation as a live demo and none of the deniability. By scripting the demo, Google removed every excuse the model would have had if it had hallucinated in real time, and they still let the error through.
The Pattern to Watch For
Any AI demo your CEO will sign their name to is a public claim about your model’s reliability. The competitive pressure to ship a response — to a competitor’s launch, to an analyst day, to a board meeting — is exactly the pressure that defeats the verification step. If your launch timeline is set by your competitor, your verification process is set by your competitor too, and that is not a process you control.
What You Should Steal
Make the rule: no AI output goes into external marketing without a named subject-matter expert signing off in writing on the factual claims. Not the AI team. Not the marketing team. The astronomer, the lawyer, the doctor — whoever owns the truth in that domain. Google’s $100B day was the cost of skipping a check that takes one expert ten minutes. The ratio of that cost to that prevention is the right thing to put on the wall in your launch room.