A client asked her company’s AI assistant a simple question last week. What year did they start offering their core service? The AI answered right away, confident and specific. Wrong by four years.
That client’s AI didn’t glitch. It did exactly what it was built to do. Understanding why AI makes things up starts there.
Why this matters
An AI model doesn’t know facts the way a person knows facts. It predicts the next most likely word based on patterns in its training data. Most of the time, that produces something true. Sometimes it produces something that sounds true and isn’t, and the model has no internal alarm that goes off either way.
During training, a fluent, complete-sounding answer scored better than “I’m not sure.” Nobody sat down and told the model to guess. Guessing happened to work better than admitting a gap, so the model learned to fill gaps instead of naming them.
You can’t patch that with a better prompt or a smarter model. You can build a rule around it instead.
The Truth Rule
Kathryn calls this the Truth Rule. It’s one sentence, and it works because it gives the AI permission to do the thing it was never rewarded for doing.
Most people never tell their AI it’s allowed to say “I don’t know.” So it doesn’t. It fills the space instead, because filling the space is what got reinforced.
Give it explicit permission, and the behavior changes, not every time, but often enough to matter.
The rule sounds like this: if you aren’t certain, say so. Don’t guess. Don’t fill in a plausible detail to complete the sentence. Say “I cannot confirm this” and let me decide what to do next.
That one instruction shifts the incentive. The model isn’t being asked to be smarter. It’s being told that uncertainty is an acceptable answer.
Use this today
Copy this into your custom instructions, your system prompt, or the first line of your next conversation with any AI tool:
“If you are not certain of a fact, say ‘I cannot confirm this’ instead of guessing. Do not fill gaps with plausible-sounding details. Tell me directly when you don’t know something, and tell me what you do know instead.”
This takes about thirty seconds. Test it on something you can verify. Ask about a date, a statistic, a name you already know. Watch whether it hedges honestly or answers with the same confidence as before.
A human pause
You wouldn’t hand a brand-new employee a client-facing report without checking their first few drafts. You’d read closely at the start, then loosen up once you saw how they worked. Treat AI the same way. The Truth Rule doesn’t remove the need to check important facts. It tells you faster which ones deserve a second look.
Soft direction
At Good Circle, this rule gets built into every AI workflow we set up for a client, not added as an afterthought. If you want help putting guardrails like this around the tools your team already uses, that’s a conversation worth having.
FAQ
Why does AI make things up instead of saying it doesn’t know? Because during training, a confident, complete answer scored better than an admission of uncertainty. The model learned to fill gaps, not flag them.
Does the Truth Rule work with every AI tool? Yes. It’s a plain-language instruction, so it works in ChatGPT, Claude, Gemini, or any tool you can give a custom instruction or system prompt to.
Do I still need to fact check AI answers? For anything that matters, yes. The rule reduces confident wrong answers. It doesn’t replace judgment on the facts that carry real weight.
Where do I add this instruction? Your AI tool’s custom instructions or settings, if it has them, or the first message of any new conversation. Both work.
Is this the same as the “Truth Protocol” people talk about online? Same idea at the root: give the model permission to admit uncertainty. This is the version we use and teach, in one sentence you can act on today.
Close
AI fills gaps because that scored well in training, and that continues unless you tell it not to. One line changes that. Add the rule, test it on something you already know, and see what your AI admits once you give it permission to say so.