AI and math: answers to common questions from business leaders
aieducation
Can AI do math? Why does it miscount letters? Is a bigger model more accurate? Plain answers for leaders using generative AI at work.
Leaders adopting generative AI often ask the same questions about numbers. Here are straightforward answers.
Can generative AI do math?
Sometimes. It handles familiar, simple sums well because it has seen the patterns often. For new or multi-step calculations, it predicts digits that look plausible rather than computing them, so errors can appear.
Why does a chatbot miscount letters or words?
Models read tokens, which are chunks of text, not individual letters. A word like "strawberry" is split into pieces, so counting the letters inside is unreliable. The same applies to word counts in a draft.
Is a bigger or newer model always more accurate?
Larger models and step-by-step reasoning improve rough number skills, but neither guarantees an exact answer. Reasoning can also add cost, which is covered in hidden reasoning tokens.
Why are small errors the real risk?
Because a near miss looks believable. A tax total of 392.97 instead of 392.79 would pass a quick glance.
What fixes this?
- Route calculations to a calculator or code.
- Add automatic checks that re-add totals and recount words.
- Make the working visible so people can review it.
Which tasks need extra care?
Money, tax, doses, compliance figures and anything you will publish or send to a client.
What should I ask a vendor?
- Does the product use code or a calculator for math?
- Does it verify totals and counts?
- Can users see how a number was produced?
For broader guidance, see AI model selection FAQs and the Responsible AI-by-Design Framework.