\n<\/p>\n
How many Ps are in Google? According to Google, there are two.<\/p>\n
There\u2019s also is also \u201cexactly 1 \u2018r\u2019 in the word \u2018poop\u2019,\u201d Google\u2019s AI Overview says, as well as two \u2018d\u2019s in the word journalism, yet spelled it: j-o-u-r-n-a-d-i-s-m. Google did at least identify that there is one P in the last name of the U.S. president, but spelled it as t-r-p-u-m.<\/p>\n
You didn\u2019t need to be a prophet to predict that Google\u2019s AI-forward Search overhaul<\/a> was going to go over poorly. We\u2019ve done this before. The first time Google added AI Overviews to Search, the feature ended up citing satirical posts from The Onion and Reddit<\/a>, advising people to eat rocks and put glue on their pizza. <\/p>\n This time around, as Google doubles down on its commitment to make generative AI the centerpiece of its 29-year-old flagship product, it\u2019s not surprising to see it stumble.<\/p>\n \u201cCounting within words has been a known challenge for LLMs, and we\u2019re working to fix this particular issue,\u201d Google told TechCrunch in an emailed statement. <\/p>\n These basic spelling errors may seem familiar. LLMs, the kind of artificial intelligence that powers chatbots and other text-generators, are not built to understand spelling. It\u2019s been a running joke for years that whenever a company unveils a new AI model, you should ask it how many \u2018r\u2019s are in the word strawberry<\/a>. These AI models \u2014 which can code an app in seconds, or solve problems that have stumped mathematicians for decades \u2014 are about as good as a kindergartener at spelling.<\/p>\n Google\u2019s AI overview woes reach beyond silly spelling mistakes though. Google already patched an issue from last week in which searching the word \u201cdisregard<\/a>\u201d would yield what looked like a dictionary definition of the word, only the definition was shown as, \u201cUnderstood. Let me know whenever you have a new prompt or question!\u201d But these spelling errors have remained amusing because they\u2019re so difficult to quash.<\/p>\n As researchers have previously explained<\/a> when we\u2019ve asked about these spelling conundrums, AI doesn\u2019t perceive sentences as units of language made up of words and letters. Many LLMs are built on transformers models, which break down text into tokens, which can be full words, syllables, or letters, depending on the model. Instead of \u201creading\u201d like a human would, the AI converts the text into numerical representations of itself, which are then contextualized to help the AI come up with a logical response.<\/p>\n \u201cLLMs are based on this transformer architecture, which notably is not actually reading text. What happens when you input a prompt is that it\u2019s translated into an encoding,\u201d Matthew Guzdial, an AI researcher and assistant professor at the University of Alberta,\u00a0told TechCrunch<\/a>. \u201cWhen it sees the word \u2018the,\u2019 it has this one encoding of what \u2018the\u2019 means, but it does not know about \u2018T,\u2019 \u2018H,\u2019 \u2018E.\u2019\u201d<\/p>\n The token-based architecture that powers LLMs like Google\u2019s AI overview is inherently limiting, and researchers haven\u2019t been optimistic that they can solve the spelling problem.<\/p>\n \u201cIt\u2019s kind of hard to get around the question of what exactly a \u2018word\u2019 should be for a language model, and even if we got human experts to agree on a perfect token vocabulary, models would probably still find it useful to \u2018chunk\u2019 things even further,\u201d Sheridan Feucht, a PhD student studying large language model interpretability at Northeastern University, told TechCrunch<\/a>. \u201cMy guess would be that there\u2019s no such thing as a perfect tokenizer due to this kind of fuzziness.\u201d<\/p>\n This isn\u2019t necessarily an urgent problem on researchers\u2019 minds, since the utility of LLMs doesn\u2019t come in their capacity to spell. But these blatant failures help us remember that AI is not perfect, even if it may sometimes seem like an all-knowing power beyond our comprehension. We cannot blindly trust AI outputs without double-checking their accuracy.<\/p>\n<\/div>\n When you purchase through links in our articles, we may earn a small commission<\/a>. This doesn\u2019t affect our editorial independence.<\/em><\/p>\n