Why ChatGPT Is Bad at Math: How LLMs Treat Numbers as Text

When AI Can’t Count to 100

AI is being promised to surpass human intelligence across the board — and yet it still can’t reliably handle basic math. The gap between the hype and reality becomes obvious the moment you ask it to do something simple.

A few days ago, I asked ChatGPT’s voice assistant to count to 50. It made two mistakes. I tried again, counting to 100 — more mistakes. ChatGPT skips numbers, has a particular aversion to 47 and 87, and once even forgot 10.

LLM Math Errors Are More Common Than You Think

If you search online, you’ll find plenty of examples: mistakes in basic arithmetic, simple word problems, counting the number of words in a sentence, logical mathematical reasoning, and more. These aren’t edge cases — they’re a consistent pattern.

Why LLMs Struggle with Numbers

The root cause comes down to how large language models work. Like all LLMs, ChatGPT predicts the next token based on patterns it has seen before. For an LLM, math is text — not numerical computation.

At the core of the model’s “thinking” are textual patterns, not logical ones. These models are trained on massive volumes of written language, not on mathematical databases or structured numerical reasoning systems.

LLM Math Limitations: Where Things Stand Today

So for now, trust is limited — but there’s hope for what comes next. Researchers are actively working on improving mathematical reasoning in LLMs, whether through tool use, fine-tuning on formal math, or hybrid approaches that offload computation to dedicated solvers.

Have you ever caught an AI making a math error? Share your examples in the comments.