Give AI permission to say "I don't know"Give AI permission to say "I don't know"
Give AI permission to say "I don't know" to reduce hallucinations and obtain reliable results.
Give AI Permission to Say "I Don't Know"
Many AI responses are not incorrect because the model is poor.
They are incorrect because the model must not stop.
A language model is trained to respond.
Not to remain silent at the right moment.
When information is missing, almost the same thing always happens:
- the AI fills the gaps
- it sounds convincing
- it misses the mark
If you want reliable results, you need a simple but extremely effective lever:
Allow the model to say "I don't know".
Why AI Would Rather Invent Than Remain Silent
LLMs are trained to continue texts.
An empty output is considered a failure from the model's perspective.
A plausible text, on the other hand, is a success.
When information is missing, the model has two options:
The ultimate shortcut to
flawless AI results
Stop wasting time guessing prompts. Get consistent, professional AI results right from the first try, every time.
2. deliver something plausible
Without explicit instructions, it almost always chooses option 2.
The Misconception: "The Model Will Realize When It Doesn't Know Something"
A language model does not realize that it does not know something.
It only recognizes that information is missing.
And it is precisely this gap that it attempts to fill in a statistically plausible manner.
For you, this appears to be hallucination.
For the model, it is expected, correct behavior.
The Most Important Sentence Against Invented Answers
There is a phrase that immediately increases reliability:
If a piece of information is not contained in the context, explicitly say:
- it takes the pressure off the model to invent something
- it defines a legitimate null response
- it shifts the focus from completeness to correctness
In short: The model is allowed to stop.
A Practical Example
"Explain the cause of the incident and describe the measures."
- plausible but invented causes
- well-sounding but unconfirmed measures
- If information is missing or unconfirmed, explicitly say "I don't know".
- Cause: still unclear
- Measures: investigation is ongoing
- clear demarcation
- no invented details
- higher credibility
Interplay with Output Requirements
The permission to remain silent works best in combination with clear output specifications.
- List confirmed facts
- List open points separately
- No assumptions
- If information is missing: "I don't know"
Why output is important here is explained here:
Output Requirements: The Underestimated Superpower for Consistent Results (coming soon).
Why This Should Not Be an Exception but the Standard
Many treat this rule like an exception.
In reality, it should always be part of a prompt when:
- the facts are incomplete
- there are legal or communicative risks
- decisions are based on them
"I don't know" is not an error.
The Transition to Systematization
If you use this rule regularly, you will notice:
- it is forgotten
- it is easily formulated differently
- it suddenly disappears
Such rules do not belong in the head,
Why this step is worthwhile can be read here:
When a Prompt Library Really Becomes Worthwhile.
Conclusion
Hallucinations cannot be completely prevented.
But they can be massively reduced.
Not through stricter models.
Not through longer prompts.
But through a simple permission:
You are allowed to say "I don't know".
This does not make AI weaker.
