Google's AI Invented a Cheese Statistic and Put It in a Super Bowl Ad

What happened

The Super Bowl is one of the most expensive and closely watched advertising slots in the American calendar, and in February 2025 Google used a local Wisconsin placement to demonstrate what its AI assistant could do for small businesses. The ad showed a Wisconsin cheesemaker asking Gemini to help write a product description for a Gouda product. Gemini produced one, and embedded in it a statistic that does not exist: that Gouda accounts for "50 to 60 percent of global cheese consumption."

Wisconsin is the highest cheese-producing state in the US, which explains the targeting. What it does not explain is the number. Experts in the dairy industry say no reliable data places Gouda anywhere close to that level of global share. Gouda is popular, particularly in Europe, but claiming it represents the majority of cheese consumed worldwide is not an interpretation of ambiguous data. It is a fabrication with no sourcing behind it.

Google's initial response made the situation worse. Jerry Dischler, a cloud executive at the company, defended the output publicly, insisting Gemini was "functioning properly" and that the statistic was "grounded in the web," meaning the model had encountered the figure on multiple sites before surfacing it in the ad copy. That framing is technically accurate about how the model works and entirely wrong as a defense. A number that circulates across unreliable web pages is not a verified fact. Gemini repeated the figure; it did not check it.

Google eventually edited the ad and removed the claim. But the episode was already visible to a national audience and had directly undermined the message the ad was built to deliver: that Gemini is a trustworthy writing partner for small business owners. A fabricated statistic placed inside a product description, delivered as confident polished output, is precisely the failure mode that erodes confidence in AI-assisted work at the moment of its widest showcase.

The deeper problem is not that the model hallucinated. It is that the hallucination made it to broadcast. The copy moved from model output to a finalized ad without a review step that would have flagged a claim no external source could confirm. There is no record of who approved the statistic, who checked it against any industry data, or what standard the output was held to before it aired. A provable record of what a system produced and what human sign-off it passed through before going live would have caught this before it reached an audience. Without that record, the only available correction was public embarrassment after the fact.

Reported impact

Affected parties

Not publicly disclosed

Harm type

Not publicly disclosed

Scale

Not publicly disclosed

Financial impact

Not publicly disclosed

Regulatory action

Not publicly disclosed

Classification

Organization

Not publicly disclosed

AI system

Not publicly disclosed

Industry

Not publicly disclosed

Country

Not publicly disclosed

Provider

Not publicly disclosed

Incident type

Not publicly disclosed

Relevant governance controls

Governance control mapping is not available for this record.

• No controls mappedNot publicly disclosed

Control mapping is analytical. It does not state that any control would have prevented the incident.

Sources and evidence

This record was researched and written by the Index. The event is also catalogued in the following database, which is listed for cross-reference.