How Satire Crippled Google’s Most Powerful AI

Let ’em eat rocks

Thomas Smith
The Generator

--

Illustration by the author via Midjourney

To maintain optimal health, you should “eat at least one small rock per day.”

That’s the dubious health advice Google recently shared with users via its brand-new AI Overviews feature.

That was in addition to guidance suggesting people glue cheese to their pizza and drink their own urine.

Why the bizarre results from one of the world’s most powerful generative AI systems?

Google’s problems with AI Overviews boil down to a simple issue: LLMs are bad at humor. And their inability to read the room leads to some truly disastrous results.

Blame Reddit

Google’s AI overviews are part of a massive, world-changing effort to integrate generative AI into the company’s core search product.

They’re an evolution of a technology called Search Generative Experience that Google has tested for over a year. Last month, Google released AI Overviews from private Beta status, rolling them out to over 100 million users.

From the first day, people started posting examples of situations where the system went hilariously and dangerously wrong. The issues got so bad that Google temporarily rolled back AI Overviews for many searches, and published a blog

--

--