Not reliably once the numbers get long enough for a wrong digit to look just as plausible as the right one.
Short answer
Not reliably beyond short, simple sums. A language model predicts a plausible-looking answer rather than running a calculation, so accuracy degrades as numbers get longer. The more useful business question is what Canadian courts already require of anyone relying on AI output for something that matters: verify it, because responsibility does not shift to the tool.
A model is predicting a plausible-looking sequence of digits, not executing an algorithm the way a calculator does — so a longer number simply means more places for the prediction to drift from the true one. The full mechanism, including why this cannot be trained away, is covered in why AI can’t do arithmetic reliably. This page focuses on the more practical half of the question: given that, what should a business actually do about it.
Ontario’s Superior Court of Justice has put this in writing for AI use generally: “it is the responsibility of all counsel and litigants to guarantee accuracy when preparing materials for use in court proceedings, and particularly when using AI, regardless of whether they directly interacted with the technology… The court will not tolerate inadvertence in this regard.” (Consolidated Civil Provincial Practice Direction)
That rule was written for fabricated legal citations, but the underlying principle transfers directly to a wrong AI-generated number a business relies on — an invoice total, a payroll figure, a quoted price. Responsibility for checking it sits with the person relying on it, not with the tool that produced it, and that is true whether or not anyone at the business ever saw the calculation happen.
More training data on arithmetic examples helps only at the margins, because it does not remove the structural mismatch between predicting a digit sequence and computing one. The products that get this reliably right hand the calculation off to something that actually computes, rather than asking the model to predict the digits — see why AI can’t do arithmetic reliably for exactly how that works in practice. Until you know a specific product does that, treat any AI-generated number the same way the courts now treat an AI-generated citation: check it before it goes anywhere that matters.
See how ongoing accuracy actually gets checked once an AI system is in production.
