Why does ChatGPT make up numbers and how can you check them fast?

Q: Why does ChatGPT make up numbers and how can you check them fast?

ChatGPT often invents numbers because it is a word prediction engine, not a calculator or a database of facts. Language models predict which word makes the most logical sense next. This allows them to write fictitious numbers very convincingly. For serious work, always treat AI-generated numbers as drafts requiring verification.

This is the number one fear of office workers who want to use AI for serious tasks. This exact question lands in my inbox constantly.

You ask for market research. The AI spits out market share data complete with percentages. You copy and paste it into a presentation, show it to the board, and the numbers are fake. Your career could end that same day. This type of hallucination is extremely dangerous because the AI writes beautifully. Its structure looks logical and highly convincing.

Saving companies money is an old habit of mine, from well before AI became a trend. As a product leader, I drove more than 4 million US dollars in savings per year through various initiatives. These included process improvements, cost optimization, and more efficient product decisions. Automation was just one piece of the puzzle. Back then, that kind of leverage required a senior position, an engineering team, and expensive systems. Today, AI is the sharpest tool for that same habit, and it can be trained into everyone on your team.

But that massive potential comes with the risk of hallucination. I am extremely strict about data validation. I push my teams to use AI for data processing, but with one hard rule: always verify the numbers.

Here are three ways to check numbers from an AI in under two minutes, before that data reaches your boss:

Demand direct source links. Do not just ask where the numbers came from. Instruct the AI to provide a direct URL to the reference. Then click that link. If the link is dead, throws an error, or the page completely lacks the numbers, throw the data away. Do not use it.
Use Perplexity for fact finding. ChatGPT is excellent for brainstorming, writing, and structuring strategy. But Perplexity is much more reliable for finding raw data. It automatically searches the web and provides clickable citations for its numbers. In the Applied-AI Certification program I created, I always train participants this way: if you need solid data for work, start with an AI designed to provide clear references.
Apply basic business sense. If the AI claims a metric jumped one thousand percent overnight without any context, your brain must reject it. Use your logic. If the numbers look too perfect to be true, a machine probably made them up.

The corporate world has zero tolerance for handing fake data to management. Make verification a core habit.

If you want to train your team to use AI safely and boost efficiency without the fear of hallucinations, you can learn with us in the AI Circle community.