Last month, I wrote about a curious incident with Google’s AI Overviews (i.e., those AI-generated summaries that Google pushes on us instead of actual search results). If you asked Google’s AI a rather silly question — “How many days of the week have a fish in them?” — it would get tripped up and return a bunch of nonsensical results, insisting that some days of the week do, in fact, contain the word “fish” in them.
Now, just to be clear, no day of the week contains the word “fish” in its name. This seemed like an obvious bug akin to Google’s recommendations that people glue cheese to their pizza and eat rocks. I assumed that Google would use some of their billions of dollars to fix this rather glaring issue. So here we are, a month later, and as you can see, Google has addressed the issue. Sort of.
Instead of returning a different wrong answer every time I pose the question, Google now just sticks with a single wrong answer. Which I guess is progress, although the wonky formatting makes me worried that my simple question gave Google’s fake brain an aneurysm.
Interestingly, when I pose this question to Gemini directly instead of going through Google’s search, it gives a more accurate answer: “If we are looking at standard calendar names — zero. None of the days of the week are named after or contain the name of a fish.” But it follows that with a curious assertion: “It is just a classic nonsense riddle meant to trip people (and algorithms) up!”
Now, to be fair, Google isn’t the only one who gets it wrong. Claude and ChatGPT also returned absurd answers to my question. But are people really tripped up by this “classic” riddle? And is it even “classic”? I’d never heard of it until last month, and I suspect that most people quickly cop to its obviously nonsensical nature. So is Google’s assertion really just intended to make it more palatable that AI chatbots, even those created by some of the world’s biggest tech companies, still get it wrong even after all this time?
On the one hand, it is the sort of humorous aside that I’d write if I wanted to make my AI chatbot a bit more endearing. On the other hand, if AI still gets something so simple and obvious wrong, what does that say about its reliability concerning more important and complicated matters?
Opus? Want to support my writing? Become a subscriber for just
$5/month or
$50/year.