Why does generative AI struggle with numbers?

Generative AI is built on language, not arithmetic. That's why a chatbot can write a polished essay and still get a tax total slightly wrong. For anyone using generative AI for business, it's worth understanding why.

Why it miscounts

Language models don't see letters or words. They see chunks called tokens, like "straw" and "berry," with the letters sealed inside. That's why a chatbot may say "strawberry" has two Rs instead of three, or return 118 words when you asked for 150.

Why math is predicted, not calculated

For familiar sums like 2 × 2, the model has seen the pattern so often that "4" is the clear winner. For a new sum, such as $347.60 plus 13% tax, there's no strong pattern to recall. The model picks digits that look right and can return 392.97 when the correct answer is 392.79.

The risk is that close looks exactly like right. A tax total, a percent change or a BMI can each be slightly off and still look reasonable.

What actually helps

• Use a calculator or code: the AI reads the question, code does the exact math, and the AI writes the result back.
• Add checks: good products recount words, re-add totals and send a draft back if it's off.
• Ask to see the working: double-check anything that matters, like money, doses or anything you'll publish.

Bigger models and step-by-step thinking improve rough number skills, but neither guarantees an exact answer.

Questions to ask before choosing an AI product

1. Does it use a calculator or code for math?
2. Does it check totals and counts?
3. Can you see how it got the number?

These are part of a broader responsible AI approach. See the Responsible AI-by-Design Framework for how Silverberry thinks about trust, or explore how AI thinks for more on model behavior.

Generative AI is great with words, so check the numbers.