Google’s AI Overviews keep mangling simple spelling questions

Google’s AI search pitch is colliding with a basic credibility problem: a product that sits at the top of search results is getting simple spelling questions wrong. Recent examples have shown AI Overviews miscounting letters in common words, including “Google” and “astronomical,” while another test prompted the system to say there is exactly 1 “r” in “poop” and to misspell “journalism” as “j-o-u-r-n-a-d-i-s-m.”

That matters because AI Overviews is no side project. Google rolled it out to everyone in the United States on May 14, 2024, after first introducing the technology as Search Generative Experience at Google I/O 2023. At launch, Google said people had already used AI Overviews billions of times in Search Labs and projected the feature would reach more than a billion users by the end of 2024. The current mistakes are landing inside a core Google Search product, not an experimental corner of it.

The public reaction has been sharp because the failures are so easy to trigger and so hard to excuse. Users have been feeding the system bait questions, such as how many letters are in “Google” or whether there is an “r” in “strawberry,” and getting responses that expose how brittle the feature can be when it is pushed beyond a comfortable pattern. Google has already patched at least one odd AI Overview behavior from the prior week, when a search for “disregard” surfaced a chatbot-style instruction instead of a dictionary-style answer.

Google chief executive Sundar Pichai has acknowledged the problem in public, saying one AI Search result was “probably more opinionated than it should be.” That line now reads less like a stray comment than a warning about how aggressively Google is pushing generative AI into search before the product has earned the same trust users once gave a plain results page.

The spelling glitches also fit a broader pattern. Google support forum users have complained that Gemini makes spelling mistakes in generated images, and Google’s own advice has included rephrasing prompts and explicitly telling the model to spell things correctly. The larger issue is that large language models are built to predict tokens, not to guarantee orthography, which leaves a gap between Google’s AI branding and the ordinary accuracy people expect when they ask a search engine how many letters are in a word. For publishers, that gap has another cost: the more AI Overviews answers questions directly, the less reason users may have to click through to source sites.