{
    "success": true,
    "data": {
        "id": 1769978,
        "msgid": "why-cant-google-ai-overview-spell-words-correctly-1780024558",
        "date": "2026-05-28 14:30:00",
        "title": "Why can't Google AI Overview spell words correctly?",
        "author": "Gana Buana",
        "source": "MEDIA_INDONESIA",
        "tags": "",
        "topic": "Technology",
        "summary": "Google's AI Overview has been found to make basic spelling errors, such as miscounting letters in common words and misspelling terms like 'journalism', despite its advanced capabilities. Experts attribute this to inherent limitations in Large Language Model architecture, not mere software flaws. The findings underscore the importance of human verification when relying on AI-generated information.",
        "content": "<p>Have you ever imagined a technology capable of cracking complex codes\nor solving advanced mathematical problems, yet failing completely when\nasked to count letters in a word? This phenomenon is now plaguing Google\nAI Overview. Despite being heralded as the future of search engines,\nthis AI exhibits a fundamental flaw \u2014 an inability to spell correctly.\nAccording to a recent report dated 28 May 2026, Google AI Overview\nprovided surprising responses to basic queries. For instance, it claimed\nthere are two \u2018P\u2019s in \u2019Google\u2019. It also stated there\u2019s only one \u2018r\u2019 in\n\u2018poop\u2019 and misspelled \u2018journalism\u2019 as \u2018j-o-u-r-n-a-d-i-s-m\u2019. Even for\npublic figures, the AI correctly identified the number of letters but\njumbled the order, such as spelling the US President\u2019s surname as\n\u2018t-r-p-u-m\u2019. These errors add to the long list of AI \u2018hallucinations\u2019,\nfollowing previous suggestions for users to put glue on pizza or eat\nstones based on satirical sources. \u2018Counting letters in words is a known\nchallenge for LLMs, and we are working to fix this specific issue,\u2019 a\nGoogle representative stated in an official statement cited by\nTechCrunch on Thursday, 28 May. This issue is not merely a software\nglitch but a limitation in the fundamental architecture of Large\nLanguage Models (LLMs). Here\u2019s the technical reason why AI struggles\nwith spelling: Matthew Guzdial, an AI researcher at the University of\nAlberta, explained that when AI processes a word, it has a single\nencoding for its meaning but does not \u2018see\u2019 the individual letters. This\nis why classic challenges like counting \u2018r\u2019s in \u2019strawberry\u2019 often fail\neven for the most advanced AI models. Despite Google\u2019s ongoing efforts\nto resolve this, experts like Sheridan Feucht of Northeastern University\ndoubt a \u2018perfect tokenisation\u2019 solution exists. This spelling failure\nserves as a crucial reminder that AI is not omniscient. Human validation\nremains key when consuming AI-generated information. Always double-check\nsensitive data provided by AI.<\/p>",
        "url": "https:\/\/jawawa.id\/newsitem\/why-cant-google-ai-overview-spell-words-correctly-1780024558",
        "image": ""
    },
    "sponsor": "Okusi Associates",
    "sponsor_url": "https:\/\/okusiassociates.com"
}