The Power of 'I Don't Know' in AI Conversations

In most professional contexts, "I don't know" is a phrase that carries risk. It can signal incompetence, unpreparedness, or the kind of uncertainty that erodes confidence. We learn, early in our careers, to reach for confident-sounding responses even when our confidence is not fully warranted.

AI has changed this calculus, in an unexpected way.

What AI makes visible

When everyone has access to AI tools that produce confident, comprehensive-sounding responses, the person who says "I don't know, but here is how I would think about finding out" stands out in a new way. They are demonstrating something that AI cannot demonstrate: intellectual humility combined with genuine curiosity.

This combination is rare and valuable. Intellectual humility, without curiosity, is passive. Curiosity, without intellectual humility, is often just confidence dressed as exploration. Together, they produce the thing that actually generates novel insight: the genuine openness to being surprised by the answer to a question you cared enough to ask honestly.

In leadership

Leaders who say "I don't know" appropriately, meaning genuinely, in response to things they genuinely do not know, create something valuable in their teams: psychological safety around uncertainty. When the leader is willing to not-know, the team becomes willing to not-know. When the team is willing to not-know, they ask better questions, raise earlier warnings, and bring more honest information upward.

The leader who maintains the performance of omniscience produces the opposite: a team that tells them what they want to hear, that filters bad news, that rounds up their confidence to match the leader's apparent certainty. In a rapidly changing environment, that information environment is genuinely dangerous.

In AI interactions

The practice of "I don't know" extends to AI interactions in a specific way. The best AI users I know are the ones who approach each interaction with genuine openness about what they might be wrong about, who use AI not to confirm what they already think but to genuinely test it.

A worked example from a real team

A product lead I worked with made a habit of saying, in weekly reviews, exactly which parts of the roadmap she was genuinely unsure about, not just the parts that were safely settled. The first few times, the room went quiet, because nobody was used to a leader naming uncertainty out loud in a meeting where everyone else was performing confidence. Within a few weeks, something shifted. Engineers started flagging technical risks earlier, before they became expensive to fix, because the norm in the room had changed from projecting certainty to naming what was actually known. The roadmap did not get worse because the leader admitted uncertainty. It got more accurate, because the information reaching her stopped being filtered through everyone else's guess about what she wanted to hear.

Why this is harder than it sounds

Saying I do not know is easy to perform and hard to actually mean. The performed version, a leader saying it while clearly signaling that the question is trivial or that they have a strong private guess anyway, does not produce psychological safety, people can tell the difference. The genuine version requires actually sitting in the discomfort of not having an answer in front of people who may judge you for it, and doing that consistently enough that it reads as a pattern rather than a one-time admission. This is uncomfortable precisely because it is supposed to be, and the discomfort is most of what makes it credible.

The distinction that makes this useful, not just humble

There is a version of not-knowing that is just avoidance, a failure to do the work of forming a view. That is not what is being described here. The useful version pairs the admission with a concrete next step: I do not know, and here is exactly how I would go about finding out, or here is what I would need to see to have a view. That pairing is what separates intellectual humility from indecision. Without the second half, I do not know is an abdication. With it, it is the opening move of genuine inquiry.

• Say what you genuinely do not know, specifically, not as a general disclaimer
• Pair it immediately with how you would go about finding out
• Notice whether your team's honesty about uncertainty increases after you model it, that is the actual signal it worked
• Apply the same standard to AI output: name what you are unsure the answer accounts for, then check it
• Resist the version that is performance rather than genuine, people notice the difference quickly

The version of this that goes wrong

Not-knowing can curdle into something else if it becomes a permanent posture rather than a temporary honest state on the way to an answer. A leader who says I do not know about everything, indefinitely, without ever converting that uncertainty into a decision, is not modeling humility, they are abdicating the actual job, which includes deciding under uncertainty when a decision is genuinely needed. The useful version of not-knowing has a shelf life. It is honest about the current state, paired with a concrete plan to reduce the uncertainty, and followed, eventually, by an actual call, made with whatever information turned out to be available in the time that mattered.

This is the distinction that keeps intellectual humility from being mistaken for indecisiveness, and it matters more, not less, in an AI-saturated environment where a confident-sounding answer is always one prompt away. The temptation, once someone has said I do not know publicly, is to quietly resolve the discomfort by grabbing the nearest fluent AI answer and presenting it as if it had settled the question. That defeats the entire purpose. The discomfort of genuinely not knowing is supposed to be sat with long enough to do real work, not outsourced the moment it becomes uncomfortable to hold.

It is worth noticing, finally, that this whole practice runs against a career instinct most people were trained into early: that certainty reads as competence and uncertainty reads as weakness. Unlearning that instinct takes longer than adopting a new habit usually does, because it means tolerating the discomfort of looking less impressive in the short term for the sake of being more accurate in the long term. The people who get good at this are not naturally more comfortable with uncertainty than anyone else. They have simply practiced tolerating the discomfort long enough that it stopped costing them as much, which is itself a kind of skill, built the same way any other skill is built.

None of this requires a dramatic personality change, and it is worth saying that plainly for anyone who reads this and assumes it describes a kind of leader they are not naturally built to be. It is a practice, made of small, repeatable moments, said out loud often enough that it becomes ordinary rather than exceptional. That ordinariness, not any grand gesture of vulnerability, is what eventually makes it credible to the people around you.

And it compounds the same way any other kind of trust does. The first time it costs something, mostly nothing. The tenth time, it has become the norm the whole team operates inside, and undoing it is far harder than building it was.

That is the actual return on saying it. Not a single moment of admirable honesty, but a slow, compounding shift in what an entire team believes is safe to say out loud, which turns out to be worth far more than the discomfort of the first time it was said.

"What am I missing in this analysis?" "What would a skeptic say about this approach?" "Where is this argument weakest?" These prompts require intellectual humility, the genuine willingness to receive an answer that challenges your position, and they produce significantly better AI outputs than prompts designed to confirm what you already believe.