An AI that says "I don't know" — and why that's the hard part
2026-08-08 · 5 min read
Language models are trained to produce plausible continuations. An honest gap and a confident invention look identical to that objective. That's the whole bug.
The fix isn't a better model. It's a rule the model can't route around.
The admission rule
When confidence lands under 50, the answer must open with the literal sentence "I don't know the answer to that, but here's what I think…" — before any reasoning, never buried after it.
Ordering matters. A hedge at the end of four confident paragraphs is not a hedge; nobody reads it as one.
Question
What will the Fed do at the next meeting?
Answer
I don't know the answer to that, but here's what I think… The decision hasn't happened, so there is no fact to verify. What exists is pricing and commentary, which is not evidence of an outcome.
Straight Edge score · Low confidence
No verifiable source can exist for a future decision. This is reasoning, labeled as such.
Would change this answer: The published decision itself, or official guidance from the committee.
"No evidence found" is a complete answer
Retrieval failure gets reported as retrieval failure. If the search returned nothing usable, you're told that, along with what source would settle it — not a smooth paragraph assembled from adjacent knowledge.
Speculative padding is banned by default: "it could be", "some people think", "it's possible that". Speculation happens only when you ask for it, and it's labeled SPECULATION.
Test it
Don't take this page's word for it. Run the same prompt on Straight Edge AI and on whatever you use now, and compare what comes back attached to the answer.
Ask it something it could get wrong
Three questions, no card, no account. Every answer comes back with its score, its sources and an exportable trace.
Ask 3 questions free