Google's recently upgraded AI Overview feature, designed to integrate more artificial intelligence into search results, is exhibiting a familiar and persistent flaw: an inability to correctly count letters within words. This issue echoes earlier challenges faced by large language models (LLMs) and has led to amusing, if concerning, inaccuracies in user-facing responses.
Users have reported that simple queries, such as asking "How many Ps are in Google," can yield incorrect answers. In one instance, the AI Overview stated, "There are 2 'p's in the word Google," before incorrectly suggesting a connection to the mathematical term 'googol.' This particular issue was observed on Chrome and, at the time of reporting, persisted on Firefox.
The AI's difficulties extend beyond the letter 'P.' When prompted with a question about the number of 'R's in the word "enigmatic," the AI confidently responded that there was only one, despite the word containing two 'R's. Bizarrely, the AI then proceeded to spell out the word character by character, accurately displaying the letters but still omitting the 'R's in its initial count.
AI researchers attribute these errors to the fundamental workings of LLMs and their reliance on tokenization. As Matthew Guzdial, an AI researcher and assistant professor at the University of Alberta, explained to Techcrunch, transformer-based models do not 'read' text in the human sense. Instead, words are translated into numerical representations called tokens. "When it sees the word 'the,' it has this one encoding of what 'the' means, but it does not know about 'T' 'H' 'E,'" Guzdial stated. This process can hinder the AI's ability to perform basic lexical analysis.
Google has acknowledged the issue, with a company representative telling Techcrunch that "counting within words had been a known challenge for LLMs, and we're working to fix this particular issue." The company's response suggests that one immediate solution might involve disabling AI Overview for queries where these counting problems are likely to arise until a fix is implemented.
These spelling and counting errors are not entirely new phenomena for AI tools. However, the prominent placement of LLM-generated responses at the top of Google Search results appears to be surfacing these issues more frequently and directly impacting the user experience. The trend of integrating AI across various platforms, while promising, continues to highlight the challenges in ensuring accuracy and reliability in these rapidly evolving technologies. Other reported AI Overview misfires include responses to queries about definitions or even inappropriate subject matter, indicating a broader set of challenges the company is facing with its AI integration.