Bard's launch demo credited the James Webb telescope with a discovery ESO made in 2004

The example Google chose to show off Bard

Sundar Pichai, chief executive of Google and Alphabet, announced Bard on February 6, 2023, in a post titled “An important next step on our AI journey.”¹ He described it as “an experimental conversational AI service, powered by LaMDA,” opened first to “trusted testers” ahead of a wider public release “in the coming weeks.”¹

The post offered one worked example of what Bard could do: an animated GIF showing Bard “helping you to explain new discoveries from NASA's James Webb Space Telescope to a 9-year-old.”¹ The same GIF ran on Google's Twitter account that day.² Among the bulleted facts Bard's answer gave the imagined nine-year-old was this sentence, exactly as it appeared:

It is a plain, checkable sentence, and it is false. By February 8, two days after Google posted it, the error was public: Engadget reported the line and noted that the first image of a planet outside the solar system belongs to the European Southern Observatory's Very Large Telescope, from 2004.²

What a telescope actually photographed in 2004

Astronomers had already published, twice, exactly when the first exoplanet was imaged and by which instrument, and neither date is 2023.

1. April 2004

   ESO detects a candidate exoplanet³

   Using the NACO adaptive-optics instrument on Yepun, one of the Very Large Telescope's four 8.2-meter units at Paranal in Chile, the European Southern Observatory detected a faint, very red point of light beside the brown dwarf 2M1207. ESO called it a giant planet candidate companion and said confirmation was still pending: “If the candidate companion of 2M1207 is really a planet, this would be the first time that a gravitationally bound exoplanet has been imaged around a star or a brown dwarf.”³

2. 30 April 2005

   ESO confirms the first exoplanet image⁴

   A year of follow-up observations showed the object was bound to 2M1207 and about five Jupiter masses, a planet rather than a background star. ESO's own words: “Our new images show convincingly that this really is a planet, the first planet that has ever been imaged outside of our solar system.”⁴

3. 1 September 2022

   JWST takes its own first exoplanet image⁵

   NASA's Webb team announced the telescope's first direct image of a world beyond the solar system, the gas giant HIP 65426 b. The same release said so plainly. “While this is not the first direct image of an exoplanet taken from space — the Hubble Space Telescope has captured direct exoplanet images previously — HIP 65426 b points the way forward for Webb's exoplanet exploration.”⁵

So Bard's demo did not just misdate an obscure fact. It handed JWST a record ESO had held, and had said it held, for eighteen years, and it did so five months after NASA had described the telescope's own much narrower first as explicitly not a first.

A stock drop the reporting won't pin on the demo

Google did not pull the demo or delay the launch. One day after Microsoft unveiled ChatGPT built into Bing, Google went ahead on February 8 with the Paris event where it planned to show Bard more widely. Reuters reported that analysts said the event lacked details on how Google would answer Microsoft's ChatGPT challenge.⁶ That same Wednesday, Alphabet's shares fell.

Reuters put the closing move at 8 percent, or $8.59 a share, to $99.05, and more than $100 billion in market value.⁶ NPR reported the shares had slid as much as 9 percent intraday, and that Microsoft's stock rose 3 percent the same week on its own ChatGPT news.⁷ No source ties the fall to Bard's error alone. It arrived in the same two days as a Google event analysts said lacked details on how it would answer Microsoft's ChatGPT challenge, and a rival's competing launch. The reporting on the drop places all three side by side rather than picking one cause.⁶

Asked about the error, a Google spokesperson tied it to process rather than to any one person: “This highlights the importance of a rigorous testing process, something that we're kicking off this week with our Trusted Tester programme.”⁶

Why the sentence came out so sure of itself

Nothing about the JWST line was a lookup that failed. A language model like the one behind Bard writes by predicting one likely word after another from patterns in its training text, and it has no built-in check on whether the finished sentence is true. The false claim about JWST was written the same way the correct bullets in that same promotional GIF were written: word by word, from what usually follows, in a launch post Google itself had published two days earlier.¹

It also did not read as a guess. Every bullet in Bard's answer, right ones and the wrong one alike, was delivered in the same flat, declarative voice.² That is not a coincidence of style. A model's stated confidence is not a reading of how often it is actually right, so a false sentence about a space telescope comes out sounding exactly as certain as a true one about the same telescope. Fluent and confident is what the writing does by default. Correct is not part of the specification.

None of this needed the demo to be careless in some unusual way. Bard's underlying technology writes fluent, confident text by default, whether or not the specific claim is true. Google aimed that technology at exactly the kind of question where a fluent wrong answer is cheapest to produce and easiest to miss before publishing: something specific and checkable, explained to a complete novice. This was not a rare malfunction. It is the ordinary behavior of the technology, and Google's own launch ad is where it happened to surface.

The failure moved from a demo to a search bar

The mechanism did not stay inside one promotional GIF. In May 2024, Google rolled AI Overviews out across Search in the United States, putting an AI-written summary above the results for many queries. Elizabeth Reid, the company's VP of Search, wrote afterward: “But some odd, inaccurate or unhelpful AI Overviews certainly did show up.”⁸ The same post also states that AI Overviews “generally don't ‘hallucinate’” the way other language-model products do, because the feature is built to check its answers against retrieved web pages.⁸ That claim is Google's own account of its own design, not an outside finding, and it sits next to the same post's admission that wrong answers reached real users at Search's scale.

The problem is not that AI answers are usually wrong. It is that fluency was never the signal to check. Bard's launch demo and an AI Overview both write in the same finished, declarative voice, whether the underlying claim is right or invented. The only way to tell the difference is the one thing the writing itself does not do: check the claim against a source that actually settles it.