Google AI Overviews Face Scrutiny Over Basic Spelling and Letter Counting Errors

What happened

Google's generative AI feature in Search, known as AI Overviews, has been observed making elementary spelling and counting mistakes. Examples include the AI claiming there are two 'p's in the word 'Google' and misidentifying the count of letters in 'journalism' while spelling it as 'j-o-u-r-n-a-d-i-s-m.' The AI also struggled with names, identifying the correct letter count in a public figure's name but spelling the name incorrectly as 't-r-p-u-m.'

What's new in this update

In a statement to TechCrunch, Google confirmed that counting within words is a known challenge for large language models. The company indicated it is working to fix these specific issues as it continues to roll out generative AI features to its search engine. This acknowledgment comes as Google doubles down on making AI the centerpiece of its search product, despite recurring accuracy concerns.

Key details

The errors extend beyond simple counting. In one instance, a search for the word 'disregard' prompted a dictionary-style response that actually mirrored a chatbot's refusal message: 'Understood. Let me know whenever you have a new prompt or question!' These types of hallucinations are tied to the transformer architecture of LLMs. Instead of reading text character by character, the models use 'tokens'—syllables or word fragments—which are converted into numerical representations.

Background and context

This is not the first time Google’s AI-forward search overhaul has faced public criticism. Previous iterations of AI Overviews were found citing satirical content from The Onion and Reddit, leading to dangerous suggestions such as advising users to put glue on pizza. The spelling issue, often referred to as the 'strawberry' test (where models fail to count the 'r's in the word strawberry), has been a running joke in the AI industry for years, highlighting a persistent technical hurdle.

What to watch next

Researchers remain divided on whether a 'perfect' tokenizer can ever be achieved. Experts like Matthew Guzdial from the University of Alberta and Sheridan Feucht from Northeastern University suggest that while models may improve, the inherent 'fuzziness' of tokenization makes letter-perfect accuracy difficult for LLMs. Google is expected to continue patching these specific errors as they arise, but the architectural limitation remains a core challenge for the industry.