Why AI Struggles with Counting A’s in ‘Banana’ — And Why ChatGPT Proved Me Wrong

Today, I dove into the fascinating peculiarities of artificial intelligence with a simple yet surprisingly tricky question:

“Why do AI models struggle with questions like, ‘How many A’s are in the word ‘banana’?’”

What followed was an eye-opening (and a little ironic) chat with ChatGPT. It explained its limitations, nailed my test question, and left me both entertained and impressed. Here’s how it all unfolded.

The First Question: Why the Struggle?

ChatGPT started with an honest and insightful answer about why AI often struggles with questions that seem straightforward:

1. Tokenization Issues

• Many language models operate on tokenized text, where a word like “banana” might be split into subwords or tokens. The model might not process the word as a whole, making it difficult to count specific letters.
• For instance, “banana” could be tokenized into [“ba”, “na”, “na”], complicating the count of “A”s without additional logic.

2. Lack of Character-Level Understanding

• LLMs are optimized for understanding and generating text at the word or phrase…