Google’s AI overview still can’t be spelled out, and the internet knows a lot about it

Google AI Overviews decorated more than Ken Kesey when it was first introduced in 2024, creating facts about drizzling “glue on pizza,” among other charges.

And while the powerful technology at Gemini has improved its accuracy significantly over the past two years (unfortunately for publishers), AI Overview still gets basic questions wrong. And that includes spell check.

Google’s AI tools are always a pain in the ass when it comes to spelling questions, having spent the last two years trying to answer the question “how many r’s are in the word strawberry?” wrongly. But it’s still bad. On Tuesday, X user Naomi Rohatyn tested the current ability of a large language model (LLM) to answer a spelling question.

“How many e’s are there in the word astronomical?” they asked.

“There are exactly 2 ‘e’s in the word “astronomical” (astronomicael),” answered AI Overview.

This Tweet is currently unavailable. Either it is loading or it has already been downloaded.

We tried it ourselves, and got the same answer.

It seems to apply to any word with four or more syllables:

No.
Credit: Google / Mashable

It’s wrong.
Credit: Google / Mashable

Suffice it to say, people are having another fun round of testing the weaknesses of the technology spell, sharing their findings on social media.

This Tweet is currently unavailable. Either it is loading or it has already been downloaded.

This Tweet is currently unavailable. Either it is loading or it has already been downloaded.

This Tweet is currently unavailable. Either it is loading or it has already been downloaded.

Why can’t Google AI Overviews spell correctly?

I don’t want to be all Billy Madison’s spelling bee about this, but considering that users are less likely to click on links if an AI summary appears in the results, surely the information provided in the AI ​​Overview should be accurate. But it’s complicated.

AI chatbots need specific context and details to respond as well as possible, so of course spelling words within their training data seems easy. However, things get complicated when you ask LLM to consider words letter by letter, as the model will process the text in chunks rather than individual letters (called tokenisation).

I thought I’d ask Google’s Gemini itself, to see what it had to say for itself. Protect your bad spelling!

“The short answer is I don’t look or write the way you do,” Gemini said. “When you write the word ‘apple,’ your brain processes five different letters. When I read or write a text, I see the word as a single unit called a token (a numerical representation of a word or part of a word).

“Because I process words as whole letters of meaning instead of individual letters, I don’t ‘spell’ the words in order. I know exactly what the word ‘apple’ means and how it relates to other words, but I don’t naturally focus on the fact that it contains two ‘p’s unless I’m specifically asked to break them.”

Mashable reached out to Google for comment.