Why Google’s AI Still Struggles with Spelling Despite Its Advanced Capabilities
AI can also compose code, create images, abstract research papers, and even help solve complicated scientific questions. However, some of the most sophisticated AI systems find themselves unable to perform a task that most children learn in their early childhood – that of spelling.
This has been demonstrated amusingly and revealingly by recent instances from Google’s AI-powered Search. Users noted that Google’s AI Overview had mistakenly added up the letters in basic words, misspelled common ones, and even had incorrect spellings when trying to explain them. These errors can seem amusing, but the root cause of them is a major aspect of how contemporary AI systems work.
The Latest AI Spelling Faux Pas
A few responses were recently written by Google’s AI Overview and immediately garnered online attention. Sometimes the system confused the number of letters within words and sometimes it made errors in the spelling of the words it was trying to decipher.
It might be surprising how this could be possible in the age of sophisticated AI models. The same technology, after all, writes software code, can answer complex questions, can create detailed written content, in seconds. But the spelling and letter-counting issues have long been recognized as weaknesses of large language models (LLMs).
Google has admitted the problem, saying that counting letters in words is known to be a problem and it’s looking to improve.
What are the reasons why AI has yet to master simple spelling tasks?
The issue is with the design of large language models.The issue is with large language model design.
Language is composed of letters, words, grammar and meaning and that’s how we naturally process it. A single person can easily recognize each letter when reading the word strawberry and also count the number of times a certain letter occurs.
But AI models don’t read text like that, though.
In most modern AI systems, language is processed by a series of units called tokens and the architecture used is a Transformer. A token can be a whole word, a portion of a word, a syllable or sometimes a single character.
The model interprets a word as a series of letters, rather than representing the word as a string of letters. While this will enable the AI to identify the relationships between concepts and provide relevant answers, it will not necessarily enable the AI to know what each individual letter in a word is.
This means that tasks involving exact character-level analysis can be surprisingly challenging.
The Tokenization Challenge
Tokenization is at the essence of these contemporary language models.
AI systems don’t just process each letter individually; they segment the text into meaningful chunks to process language more efficiently. This method improves the performance of such operations as conversation, translation, summarization, and content generation greatly.
But there are some limitations of tokenization when the task involves counting the exact number of letters or checking the spelling.
For instance, if the word Google is part of the model, it could be considered as one token, or a combination of multiple tokens that collectively represent the concept of Google. It knows what the word means and the context, but might not necessarily be able to track individual letter within the word.
Therefore, a model that can produce programs that are sophisticated enough to pass a very basic test like counting the number of e’s in a word can be wrong for the test of counting the number of f’s in a word.
Despite its success, the problem has not been completely solved by researchers.
For years, AI researchers have known about these limitations. In fact, a simple inquiry of the amount of occurrences of a specific letter in a word is a frequent casual challenge among AI enthusiasts.
This isn’t a quick fix software glitch that can be corrected by an overnight update. It is rooted in the architecture that enables the large language models to work in the first place.
Many experts are of the opinion that there is nothing that could be done to tokenize the data completely without including other compromises. The complexity of language made it difficult to decide on how to split words into tokens and this is still open for discussion.
Future models could be better at reasoning at the character level, but for current AI models, the basic principles of design are about grasping the meaning and providing helpful answers, rather than being about spelling accuracy.
Why These Errors Matter
While it is easy for people to overlook the importance of spelling accuracy, compared to the sophistication of other AI applications, it is a reminder of the challenges of AI.
The user experience is often with AI systems that seem to be very knowledgeable and able. These are tools that can give a detailed answer to a variety of questions, which can lead to the assumption that they always have the right answers.
But, cases of misspelled words, wrong number of letters, or false data show that AI systems can still make minor errors.
These mistakes emphasize the need for verifying AI-generated content, particularly when making critical decisions, asserting facts, executing work, or creating educational material.
A Reminder That AI Is Not Infallible
The recent spelling problems in Google’s AI Overview are one of the many realities of artificial intelligence. While these models make great advances, they are not yet fully comprehending language like humans do. They show good pattern matching, anticipating text and responding appropriately to the text, but they may not perform well on tasks involving symbolic reasoning.
With AI more and more becoming a part of search engines, workplaces, and daily life, buyers ought to take it as a useful device, not a source of information that’s all the time correct.
While technology is advancing at an astounding rate, events like these serve as a reminder that even the most sophisticated AI systems can sometimes be stumped by the simplest of tasks. In the new reality of human and artificial intelligence, it is as important as ever to think critically and to verify.
