Did you know Google’s AI struggles with spelling? It’s botched “Google” more than once!
In a recent hiccup, Google’s AI determined there are two Ps in the word “Google,” while also making curious claims about other words. For instance, it identified “poop” as having exactly one ‘r’ — and then proceeded to misspell “journalism” as j-o-u-r-n-a-d-i-s-m. The AI did manage to identify one P in President Trump’s name, but hilariously spelled it t-r-p-u-m.
As Google leans into generative AI’s role within its flagship tool, it’s no surprise that stumbling blocks are cropping up.
“Counting letters correctly has been a challenge for large language models (LLMs), and we’re aware of this issue,” a Google spokesperson mentioned in a statement.
At first glance, these spelling blunders might seem trivial. Yet they raise serious questions. LLMs — the driving force behind many modern text generators — are not inherently designed for spelling. Anecdotal humor has surrounded AI models for years, where a common challenge is to ask how many ‘r’s are in “strawberry.” Despite their prowess in other areas, these systems have the spelling accuracy of a young child.
The issues with Google’s AI are not just limited to spelling, however. Recently, the platform corrected a problem where searching for the word “disregard” led to a response that amusingly said, “Understood. Let me know whenever you have a new prompt or question!” Such ongoing inaccuracies highlight a persistent conundrum.
Experts explain that AI processes sentences differently than humans do. Typically, LLMs utilize transformer models that dissect written words into tokens — not necessarily corresponding to letters or syllables. Rather than “reading” the text as we do, these systems turn it into numerical data that they use to form coherent responses.
LLMs operate on transformer architectures, which don’t precisely ‘read’ text. When given a prompt, it’s reframed into a numerical encoding. Take the word ‘the’; the AI has one representation for it but lacks awareness of the individual letters ‘T,’ ‘H,’ and ‘E.’
The token-based construct within LLMs like Google’s poses inherent challenges, and many experts are skeptical about solving the spelling issue anytime soon.
While this is not pressing for researchers, as the real value of LLMs lies outside spelling accuracy, these continued oversights serve as reminders that AI is not infallible. Trusting AI completely without verification can be a risky endeavor.