I watched an AI tool confidently invent a statistic in front of a client once, complete with a percentage and a specific year, none of it real. The client almost used it in a pitch deck. That almost is the whole reason this issue exists, and it’s the reason I link every single number I put in this newsletter.
Tell your AI this every single time: do not make things up.
AI tools don’t lie the way people lie. That distinction matters, because it changes how you defend against it. A person who lies knows the truth and chooses to say something else. AI doesn’t know anything in that sense, it generates the most statistically likely next word, one after another, and sometimes the most likely next word is a confident, specific, completely fabricated fact. There’s no intent behind it and no internal alarm that goes off. It reads exactly as convincing as a true one, in the same tone, with the same authority, because to the tool there’s no difference between the two. That’s the actual danger. It’s not that AI lies. It’s that it can be wrong with total confidence and no tell.
This has a real name, hallucination, and it’s not a bug they’ll fully patch out next quarter. It’s a side effect of how these tools fundamentally work. They’re built to produce fluent, plausible language, and fluent plus plausible is exactly what a good fabrication looks like. The better these models get at sounding right, the better they get at sounding right when they’re wrong, too. Capability and this particular risk grow together, which is why “just wait for a better model” isn’t a real answer.
The fix is embarrassingly simple and almost nobody does it. Tell it explicitly, every time, not to state anything as fact unless it can point to where that fact came from. If it doesn’t have a real source, it should say so plainly instead of guessing convincingly. This doesn’t make the tool perfect, it’ll still slip, but it dramatically raises the odds that when it doesn’t actually know something, it tells you, instead of smoothly filling the gap with something that sounds like knowledge. I didn’t trust the smooth-talking guys in college, and I still don’t in politics, OR smooth-talking digital tools. Ha.
And then, the non-negotiable second half: you check the ones that matter anyway. Any number, name, date, or claim that’s going out under your name to customers gets a two-minute verification, click the source, confirm it says what you think it says. Not because you’re paranoid. Because your name is on it, and “the AI told me” has never once worked as an apology to a customer who caught a fake fact.
Takeaway: AI doesn’t know when it’s wrong. That’s your job to catch, every single time.
Try this: Add one line to your standard AI instructions, the ones you reuse for every project:
“Never state a statistic or fact without naming its source. If you don’t have a real one, say so plainly instead of guessing.” Use it every time, not just when it feels important, because the times it matters most are exactly the times you won’t see it coming.
Talk soon, Tiffany