This is because LLMs do not inherently understand math. They stick characters together that are likely to go together based on the content they were trained on. They’re literally just glorified autocorrect.
If you want a tool that can actually do math from natural language input, try WolframAlpha.