Google's AI search feature has faced criticism for its inability to correctly spell basic words, including the company's own name. According to the AI's Overview, there are two Ps in 'Google,' but the spelling of the word appears to be incorrect. The AI also claimed there is exactly one 'r' in 'poop' and two 'd's in 'journalism,' which it spelled as 'j-o-u-r-n-a-d-i-s-m.' While the AI correctly identified one P in the last name of the U.S. president, it spelled it as 't-r-p-u-m.' These errors highlight ongoing challenges with spelling in large language models (LLMs), which are not designed to handle such tasks.
The issues have resurfaced as Google continues to integrate generative AI into its search engine, a move that has previously led to similar problems. Google acknowledged the issue in an emailed statement, stating, "Counting within words has been a known challenge for LLMs, and we’re working to fix this particular issue." These spelling mistakes, while seemingly trivial, underscore the limitations of AI systems that rely on transformer architecture to process text. Instead of reading like humans do, AI converts text into numerical representations, which can lead to such errors.
Researchers have noted that the token-based architecture of LLMs inherently limits their ability to accurately spell words. "LLMs are based on this transformer architecture, which notably is not actually reading text," explained Matthew Guzdial, an AI researcher. "When it sees the word 'the,' it has this one encoding of what 'the' means, but it does not know about 'T,' 'H,' 'E.'" The challenges with spelling in AI models remain a topic of discussion among researchers, who believe that the fundamental issue lies in how AI processes language.
"It’s kind of hard to get around the question of what exactly a 'word' should be for a language model," said Sheridan Feucht, a PhD student studying large language model interpretability. "My guess would be that there’s no such thing as a perfect tokenizer due to this kind of fuzziness." While these errors may not be urgent, they serve as a reminder that AI is not infallible. Users are advised to verify AI outputs for accuracy.
Source: techcrunch
