What happened
On 6 February 2023 Google announced Bard, its answer to ChatGPT, in a blog post and a short promotional ad posted to the company’s own Twitter account. The ad showed Bard answering the question, “What new discoveries from the James Webb Space Telescope can I tell my 9-year-old about?” One of the bullet points Bard returned said the telescope “took the very first pictures of a planet outside of our own solar system.”
That was wrong. The first image of an exoplanet was captured in 2004 by the European Southern Observatory’s Very Large Telescope, nearly two decades before James Webb launched. The astrophysicist Grant Tremblay flagged the error on Twitter on 7 February, noting that a routine Google search returned the correct answer. Reuters reported it the next morning, 8 February, the day of a Google AI event in Paris.
The detail that matters for the record is where the claim appeared. Bard had not been released to anyone; the mistake was in the marketing for the product, not in a live demonstration a journalist had prompted. Google had chosen the example, and the false claim still shipped. Alphabet’s shares fell as much as nine percent intraday and closed down about 7.7 percent that day, erasing roughly US$100 billion in market value, in the middle of the first competitive rush between Google and a newly AI-powered Microsoft Bing. Google said the episode underscored the need for a rigorous testing process before launch.
What an auditable version would have shown
The question Google could not answer cleanly in the moment was a simple one: before that bullet point went into an advertisement, was it ever checked against a source? Nothing in the workflow was designed to answer it. An auditable version records, for each factual assertion an assistant produces, whether it was checked against a trusted reference, which reference, and the result, captured at the time. With that record the telescope claim would have carried its own provenance, verified or not, and against what. An unverified assertion would have been visible as unverified before it became the centrepiece of a launch, rather than after the market had priced in the mistake.
Where the gap was
The gap was that a public factual claim, produced by a generative model, reached the world without a step that confirmed it against an authority. Generative models return fluent, confident text whether or not it is true, and the confidence is identical in both cases. A model is not the place to check a model. The control is a VerificationGate: a factual claim the assistant proposes is routed to a trusted source rather than accepted on the model’s say-so, and a claim that cannot be confirmed is held back before it ships. A ConductRecord preserves the claim, the source check, and the outcome, so the organisation can show what was verified and what was not.
What governance should have looked like
The lesson is not that Google should not have built an AI assistant. It is that a single factual claim, in a setting millions would see, needed one verification step it did not get. A claim worth putting in an advertisement is a claim worth checking against a primary source, and the check should leave a record. The cost of the missing step was not abstract: it was a same-day fall in market value larger than the annual revenue of most companies, caused by one unverified sentence about a telescope.
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)