How often is AI wrong? Why you should check its answers

Why AI gets things wrong

Tools like Claude, ChatGPT and Grok are very good at writing. But they don't look facts up the way a person checks a record. They predict the most likely words for your question, based on what they learned and what you gave them. Most of the time that works. When it doesn't, the answer still reads well, so the mistake is easy to miss.

• They work from averages. Ask whether you can afford something and you get an answer based on a typical month, not your slow months.
• They don't know your records. Unless you paste them in, your price list, policies and bank balance are guesses.
• Their information can be old. Prices and rules change, and the AI may not know.
• They sound sure either way. There's no change in tone between a right answer and a wrong one.

The mistakes that cost a business money

A clumsy sentence costs nothing. These do:

• A wrong number: an old rate in a quote, a price that changed, a total that doesn't add up.
• A wrong promise: a refund policy or delivery date your business never offered. In 2024 a Canadian tribunal ordered Air Canada to compensate a customer after its website chatbot gave wrong information about a refund policy.
• A decision on the average: "yes, you can afford it" when you can, on average, but not in a slow winter.
• A dropped word: "do not" becoming "do", or "only" disappearing from a condition.

Can't you just ask the AI to double-check?

You can, and sometimes it catches its own slip. But it is the same AI checking the same work, with the same blind spots. It still doesn't have your records, it still works from averages, and it can give a different answer if you ask twice.

How to check an AI answer yourself

• Check every number against the source. Prices, rates, dates and totals: open the record, don't trust memory.
• Ask what it assumed. "What numbers did you use?" shows you where an average or an old rate slipped in.
• Ask for the bad case. "What happens in a slow month?" is often where the answer changes.
• Read the small words. Not, only, unless, at least, up to: one change flips the meaning.
• Decide how sure you need to be. A social post can be checked lightly. A quote, a payment or a hire can't.

Or let Gnomon check every time

Gnomon plugs into the AI you already use. Whenever your AI gives you an answer or does a task for you, Gnomon checks the facts against your records and rules, and runs the decision a million times to see how often it works out. It takes milliseconds, and your AI comes back with an answer that holds up.

Read how the million runs work in what is a Monte Carlo simulation?, or see it on a real decision in can I afford to hire someone?