AI Doesn't Lie to You. It Guesses.

AI doesn't lie to you. It guesses. For anyone using it to make real business decisions, that distinction matters more than most people realise.

Every AI predicts what word comes next, based on everything it's been trained on. Most of the time that's accurate enough to be useful. But every model has a setting that says "I don't know," and the moment it half-recognises what you're asking, that setting switches off. Because the model is built to be helpful, it answers anyway, even when it's working without all the facts. It's most likely to do this with specific numbers, names, dates, and sources, or anything niche or recent.

Even KPMG got caught out. Their report on AI hallucinations had AI hallucinations in it. If that's where the bar sits for a firm of that scale, it's worth taking seriously.

Three fixes I now give every client I work with. First: ask it to prove what it said. Type "what are your sources for that" and ask it to go and actually check them. Second: take the pressure off before you ask anything important. Say "it's fine if you don't know" before the question. That one phrase reduces the likelihood of the model filling in gaps with confident-sounding guesses. Third: open a fresh chat for anything that matters. Paste the answer in and ask it to verify every claim. A fresh chat checks more honestly than the one that made the claim in the first place.

AI is not wrong by accident. It's built to answer before it's built to be right. Once you understand that, you start using it on your own terms rather than trusting it on its.