A months-in-the-making launch event for Google Bard contained one falsifiable claim about the James Webb Space Telescope. Within 24 hours, Alphabet's market cap dropped by $100 billion. A second AI checking the first would have caught it.
In February 2023, Google rushed a launch event for its new AI chatbot, Bard, after watching ChatGPT detonate the public imagination two months earlier. The demo featured Bard answering simple, friendly questions, including one about new discoveries from the James Webb Space Telescope to share with a 9-year-old. Bard wrote that the telescope had "taken the very first pictures of a planet outside of our own solar system." The claim was false. The first exoplanet image was captured in 2004 by the European Southern Observatory's Very Large Telescope, almost two decades before JWST launched. Within twenty-four hours of the error being noticed, Alphabet's share price fell 7.7%, erasing about $100 billion in market value. One wrong sentence, in a launch event Google had spent months preparing.
The error was first flagged on Twitter by astrophysicist Grant Tremblay, who pointed out the factual mistake within hours of the demo. Reuters picked up the story. The market's reaction was visceral. Investors had been waiting to see whether Google could respond to ChatGPT with something credible, and the very first public-facing example contained an obvious, falsifiable mistake about a topic the AI was specifically being marketed as authoritative on. By the close of trading the next day, shares had lost roughly $100 billion in value.
What's striking is that the error was the kind of thing a college student writing a research paper would catch with a single Google search. The screenshot of Bard's answer had been live on Google's official Twitter feed for hours before anyone in the company seems to have verified it. A second pair of eyes, even an automated one, would have caught it. There wasn't one.
Three years later, the demo error is still one of the most-cited examples in AI safety literature. It's a clean case study because the cost is unambiguous (a market cap movement), the error is unambiguous (a falsifiable factual claim), and the prevention is unambiguous: somebody, or something, should have asked "are you sure?" before the answer went public.
A single large language model has no way to know that it's wrong. Internally, all its outputs feel equally fluent and equally well-formed. The model that confidently said "JWST took the first pictures of an exoplanet" is the same model that would have, in a different context, confidently said the opposite. Whether the output is correct or incorrect doesn't change how the model produces it.
This is why every safety-conscious AI deployment now includes some form of verification layer: RAG to ground claims in retrieved documents, fact-checking pipelines, or, increasingly, a second LLM asked to double-check the first. The instinct is right: a single model is too prone to confident error to be deployed alone on tasks where errors are costly.
In AI analytics, the same dynamic applies. A model that produces an insight ("your revenue grew 23% week-over-week, driven by paid social") has no internal flag for "I'm confident in this" versus "I'm guessing." Both feel identical from the inside. And in analytics, errors compound: a leader who acts on a hallucinated number reallocates spend, fires or hires staff, doubles down on a campaign that wasn't actually working. The cost isn't a stock-market headline, but it's real money over real quarters.
Askalytics runs every insight past a structurally different second model before it ships to you.
The Bard demo failed because there was no second voice in the room asking are we sure? The fix is to put one there, and not a copy of the same model, but a structurally different one, so it has different failure modes.
Askalytics runs a Confidence Check at the end of every AI Insight. Claude, from Anthropic, queries your data and writes the insight, with every number tied to the query that produced it. The Confidence Check then asks a second, different AI model to review the same data and answer and assess the interpretation. If the two models disagree about whether the evidence supports the conclusion, the insight is flagged with a lower confidence score and the conflict is shown to the user. If they agree, the score reflects that consensus.
This is structurally different from a single-model self-check, where you ask the same model "are you sure about that?" Single-model self-checks are notoriously weak: the model has the same blind spots in both passes, so it tends to agree with itself. Two different models, each reading the same evidence on its own, are less likely to fail in the same place. When both independently confirm an interpretation, the user has genuine signal that the conclusion is robust. When they disagree, the user has equally valuable signal that the data is ambiguous and the recommendation is soft.
For the Bard demo, a Confidence Check would have meant a second model reading the same prompt and noticing the JWST claim was disputed by readily-available facts. The first model's answer would have been flagged. The demo would have been rewritten before going public. $100 billion in market value would not have moved.
Google's $100 billion telephone-game moment is the most-quoted AI failure in business history because it's so cleanly diagnosable. One model, one output, no verification, public stage. Three years on, the lesson has been absorbed by the AI-safety research community but is only just being built into product. If you're choosing an AI analytics tool in 2026, the question to ask is not "how accurate is your model?" It's "what does your system do when your model is wrong?"
To go deeper into how Askalytics handles disagreement between the two models, read about our Confidence Check.