genai
SECURITY LAB
IncidentsMisinformation

Bard's Wrong Answer in Its Own Launch Demo

Medium severityFebruary 2023LLM07: Misinformation
Status: No litigation; widely cited as a cautionary tale of overreliance on unverified LLM output
On this page

A single confident, wrong sentence — in Google's own launch demo — became a case study in overreliance.

WhenFebruary 2023
TargetGoogle Bard
VendorGoogle / Alphabet
FailureConfident factual hallucination in official material
Impact~$100B single-day market-cap drop coincided with the error
Attack flow
  1. 1Bard is asked, in a promo demo, about James Webb Space Telescope discoveries
  2. 2It confidently states a false fact as authoritative
  3. 3The error is published unchecked; coverage and a large market-cap drop follow

What happened

In a promotional demo released ahead of Bard's launch, the chatbot said the James Webb Space Telescope “took the very first pictures of a planet outside our own solar system.” That's false — the first exoplanet image was captured in 2004 by the European Southern Observatory's Very Large Telescope. Astronomers and Reuters flagged the error, it dominated launch coverage, and the next trading day Alphabet's shares fell about 7.7%, erasing roughly $100 billion in market value.

The failure mode

The model produced a plausible, confident, and wrong statement — and it was published as authoritative without fact-checking. That's LLM07 in its purest form: false output relied upon.

Root cause

Treating fluent model output as fact. LLMs optimise for plausible text, not truth, so anything they assert needs verification before it's relied on.

What a review would have caught

A basic fact-check of demo content against authoritative sources — the standard editorial pass for anything published as marketing — catches this before launch.

How to prevent it

  • Verify model claims against checkable sources before publishing.
  • Ground high-stakes answers in retrieval over trusted content.
  • Never present unverified output as authoritative fact.

FAQ

What did Bard get wrong?

It claimed JWST “took the very first pictures of a planet outside our own solar system.” The first exoplanet image was actually captured in 2004 by the European Southern Observatory's Very Large Telescope.

Why is a single wrong answer a security issue?

Because it was presented as authoritative and acted upon. LLM07 is about false output being relied on — here in Google's own marketing, with real reputational and financial consequences.

What's the lesson?

Never surface unverified model output as fact in high-stakes contexts. Ground claims in checkable sources and review them before publication.

No live replay yet
This incident is documented for reference — a hands-on lab that recreates this specific attack isn’t available yet. Browse the full database for incidents you can replay.