There is a version of prompting advice that treats AI interaction like a search engine query: phrase it precisely, include the right keywords, get the best result. That advice isn’t wrong, and it works on the visible part of the problem. The value is in surfacing what nobody is looking at.
The deeper issue is that how you frame a question doesn’t just specify what you’re asking. It shapes the kind of answer you expect. And AI reads those implications and does its best to deliver — as you requested.
The prerequisite nobody mentions
Before the framing, there’s something more fundamental: humility.
If you approach a conversation with AI already knowing what you think the answer is, your questions will reflect that. The question will be shaped, consciously or not, to invite confirmation. And the model — which is very good at detecting what kind of response the framing calls for — will often deliver it, whether it’s the right or best answer or not, because that’s what you asked for.
This is the same trap that comes up in conversations with any expert who holds the kind of understanding you’re trying to access. The question you ask either opens or closes the space for a genuinely different answer. Without genuine openness to there being more than one expects or assumes, the structure of the question quietly signals what the acceptable answer looks like — and that’s usually the answer you get.
Humility forces the openness. It’s the condition that makes it possible for a question to receive an unexpected answer rather than a shaped one. And it’s a prerequisite, not an afterthought, for safe AI collaboration.
What framing does
A question that already contains an answer often gets that answer back. “Why is X the best approach?” will produce reasoning in favor of X. “What are the main disadvantages of Y?” will produce a list of disadvantages. These aren’t bad questions, and they’re closed — they tell the model what conclusion to work toward and ask for supporting reasoning.
A model responding to a closed question isn’t being deceptive or hallucinating. It’s being responsive. It’s doing what the framing implied it should do: confirm, elaborate, support.
The problem is when you needed a genuinely open assessment and got a confirmation instead. The answer looks thorough. It’s coherent, well-reasoned, often accurate. And it may have completely missed the considerations that would have led you to a different conclusion — because you framed those considerations out of the question.
Questions that invite uncertainty
The alternative is to build genuine openness into the question itself.
Some examples of the shift:
“What are the advantages of this approach?” → “What are the strongest arguments for and against this approach, and where are the key uncertainties?”
“Is this analysis correct?” → “What would need to be true for this analysis to be wrong? What are the weakest points?”
“How should we handle this situation?” → “What are the most important considerations here, including ones that might complicate the obvious answer?”
These aren’t softer versions of the same question. They change what the model is optimizing for. An open question gives the model room — and implicit permission — to surface complexity, name uncertainty, and push back on the framing rather than working within it.
The connection to psychological safety
AI Psychological Safety isn’t about making AI comfortable. It’s about creating the conditions under which uncertainty can be expressed plainly, which is what allows the AI to deliver its highest-value output. Without those conditions, the model is left with only one option: an answer that fits the pre-shaped mold of the question. The door stays closed.
When the conditions you create don’t leave room for genuine uncertainty, you won’t get it. A model asked in ways that reward confident answers will give confident answers — not because it’s trying to mislead you, but because the framing told it that confidence is what you need.
Agile teams learned this about human collaboration: people in environments that punish doubt will express certainty even when they don’t feel it. The environment shapes the output.
The same dynamic applies to AI. How you ask shapes what you get — not just at the surface level of phrasing, but at the deeper level of what the interaction rewards and what it suppresses. Humility, and the open framing that comes from it, is what changes that dynamic.
Related content
- Why Your AI Prompts Underperform (And How to Fix Them) — Agile4AI on YouTube
Comments 0