What happened
It was reported that in February 2023, Google was preparing to introduce Bard, its answer to ChatGPT, when a factual mistake in a promotional advertisement attracted worldwide attention and contributed to a sharp fall in the company’s share price.
The advertisement showed Bard answering a question about discoveries made by the James Webb Space Telescope, including a claim that it had taken the first photographs of a planet outside our solar system. The answer sounded convincing, but it was wrong because the first such image had been captured in 2004, almost two decades before the James Webb telescope was launched.
The mistake was spotted by an astrophysicist and quickly circulated online, attracting particular attention because Google was competing with Microsoft to establish itself as a leader in generative AI.
On 8 February, Alphabet’s shares fell approximately 7.7%, wiping around 100 billion USD from its market value amid wider investor concerns about Google’s AI strategy. Although the inaccurate advertisement was not the only factor behind the decline, it became a highly visible example of the risks associated with presenting AI-generated information as fact.
What made the incident unusual was that the error appeared in Google’s own promotional material rather than in a conversation with an ordinary user, suggesting that a basic factual mistake had passed through the company’s preparation and review process before being published.
What an auditable version would have shown
An auditable system would have recorded whether Bard’s claim about the telescope had been checked against a reliable scientific source before it appeared in the advertisement, including what evidence supported the answer and whether anyone had approved it for publication.
That record would have made it possible to identify whether the mistake came from the model generating an incorrect answer or from the human review process failing to recognise that the information needed checking.
Where the gap was
Headlights identifies overconfidence and a rules-behaviour gap, where people place too much trust in information generated by a system without ensuring it meets the standards expected of the organisation using it.
Google had the expertise and resources to verify a straightforward scientific claim, yet the incorrect answer still appeared in material intended to demonstrate the capabilities of its new technology.
There is a familiar human behaviour behind this failure because information that sounds confident and comes from a sophisticated system can feel more reliable than it actually is, particularly when people are focused on demonstrating what the technology can do rather than questioning where it might be wrong.
The incident illustrates how easily a convincing answer can move from an AI system into a public statement when the people responsible for reviewing it assume that someone else has already checked the facts.
What governance should have looked like
Any organisation publishing AI-generated factual claims should have a clear process for checking important information against independent, reliable sources, particularly when those claims appear in advertising or other high-profile communications.
A VerificationGate could require factual statements to be checked before publication, while a ConductRecord could preserve evidence of which claims were reviewed, the sources used and whether verification was successful.
The wider lesson is that the quality of an AI-generated answer cannot be judged by how convincing it sounds, and even the largest technology companies remain responsible for checking information before presenting it to the public as fact.
Failure Pattern: Google’s Bard generated an incorrect scientific claim that appeared in a promotional advertisement, exposing a failure to verify AI-generated information before publication.
Governance Principle: Organisations should independently verify factual claims generated by AI and maintain records of those checks, particularly when information is being published under the organisation’s name.
The reference implementation of VerificationGate and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.
Sources
- Google AI chatbot Bard offers inaccurate information in company ad (Reuters, 8 February 2023)
- Google shares lose $100 billion after company’s AI chatbot makes an error during demo (CNN Business)
- Google shares drop $100 billion after its new AI chatbot makes a mistake (NPR)