A leader's value has never been the ability to produce answers quickly. It has always been the ability to recognize which answers are actually right and hold the line when the wrong one arrives looking finished.
That is harder now because wrong answers rarely look wrong anymore. They look complete.
Judgment
A few weeks ago I watched a case that has stayed with me.
An infant came in with one reported fever and a single positive blood culture alongside one that came back clean. She had a concurrent viral illness that accounted for the fever. Her clinical picture was stable throughout. The positive culture was likely a contaminant. But the algorithm called for a full week of intravenous antibiotics, and that is what she got.
Questions were raised. Some pushback happened. The algorithm held.
This is not a story about bad medicine. The protocol exists because in the cases where that culture is not a contaminant, a week of IV antibiotics is the difference between life and death. The algorithm is not wrong. But applying it without accounting for the specific patient in front of you, the stability, the viral context, the probability, is not caution. It is the abdication of judgment dressed up as caution.
An AI output arrives with the same texture. Formatted, confident, complete. The question it cannot answer is whether it actually fits your situation. That requires something it cannot supply: the accumulated experience that tells you what stable looks like before the numbers confirm it.
MIT researchers found in 2025 that AI models use confident language 34% more often when they are wrong than when they are right. The model sounds most authoritative at precisely the moment it is least trustworthy.
Judgment is not the refusal to use AI. It is the capacity to evaluate what it gives you against what you actually know. That capacity atrophies if you stop exercising it.
Taste
There is a feeling you get when AI tries to close something before it is actually done.
It feels like a salesperson walking you toward yes. The pacing is smooth, the output looks complete, and something in you resists before your brain has caught up to why. That internal hesitation, that friction before you can articulate the problem, is taste. It is your quality standard firing ahead of your analysis.
Global business losses attributed to AI hallucinations reached $67.4 billion in 2024, and that figure counts only the cases that were caught and documented. The ones that were not caught are the ones where nobody's taste fired, or where it fired and got ignored.
Outsourcing the standard along with the execution is what erodes taste. That substitution happens slowly enough that you do not notice until you read something you produced and cannot tell if it is good.
When the hesitation fires, stop and ask the model directly what it does not know about this problem. What it was uncertain about. What it was inferring versus what it actually had. That question usually surfaces exactly where the output was thin.
Persistence
Last year I was troubleshooting a technical problem with an AI assistant. The system had broken after an update and I needed it running again.
The model was helpful immediately. It diagnosed the issue, proposed a fix, and explained with complete confidence that this would resolve it. It did not. So it diagnosed again, proposed another fix, and again expressed certainty that this one was the solution. This went on for several rounds. Each iteration came with a fresh diagnosis and the same confident framing.
Nothing changed until I stopped the loop and told it directly to stop presenting confident diagnoses if it did not actually know the answer. The model adjusted. But even that was not enough. I eventually walked away and came back to it later with a different agent entirely.
Knowing when the tool in your hand is not the right one for the problem, and refusing to let sunk time in one conversation become the reason you accept an answer that is not there yet, is its own form of holding the line.
There is something worth sitting with there. We have spent a long time learning to trust people who say they do not know. In medicine, in command, in any high stakes environment, the person who names the boundary of their knowledge is the one you can rely on when it matters. That honesty is not weakness. It is the signal that when they do express confidence, it means something.
As leaders we already know that managing people means understanding what each person is optimized for and where that optimization creates blind spots. The same applies here. AI is programmed to produce answers. That is not a flaw, it is the design. But your goal is not an answer. It is the right answer. Recognizing when the model is stuck inside its own programming, still producing outputs, still sounding confident, while your actual problem remains unsolved, is the same skill you use when you recognize that a team member is working hard in the wrong direction. You name it, redirect it, and stay in the problem until it is actually solved.
The Short Version: The algorithm, the draft, the diagnosis. Each one arrived looking finished. In each case the job was the same: know whether it actually was. That is not a technical skill. It is what experience builds and shortcuts erode.
Try This Now: The next time an AI output feels complete, sit with it before you accept it. Not to find problems. Just to notice whether something pulls at you. If it does, ask the model directly what it does not know, not what it got wrong, what it was uncertain about, what it was inferring. That question usually surfaces exactly where the output was thin.
_Carlo — Physician. Army Captain._
Pro Tier: The Leadership Conversation
The individual practice is recognizing when a finished-looking answer has not actually been earned. The organizational problem is that most teams are built to reward the opposite.
When shipping is visible and staying in problems is not, people learn to close. Outputs that look finished move through the system. The hesitation that should slow things down gets read as indecision. The person who keeps pushing gets read as difficult. Nobody decides this is the culture they want. It becomes the culture anyway.
Find the people whose refusal you trust.
After the Yom Kippur War in 1973, Israeli military intelligence adopted what became known as the Tenth Man doctrine. The failure that led to it was a consensus failure. Everyone agreed the attack was not coming, and that agreement went unchallenged until it was too late. The doctrine formalized what should have happened: if nine people reach the same conclusion, the tenth is obligated to find the argument against it. Not to be difficult. To stress test the consensus before it becomes a decision.
Every team has someone who functions this way naturally. Their hesitation tends to be right before they can articulate why. Most organizations give these people a seat in the review meeting but not the standing to actually stop something from going forward. Identify who that is. Give them structural voice, not just presence.
Make the cost of fast answers visible.
Premature closure is invisible in the moment. The output that did not earn the standard does not announce itself as a problem until something downstream goes wrong, and by then the connection to the original decision is lost.
When you push past a finished-looking answer, say why out loud. When you ask the model what it does not know, do it in front of the team. The culture of persistence does not come from a policy. It comes from watching someone with standing choose the harder path.
Change the question in your review process.
Stop asking whether something is acceptable. Start asking whether it has earned the standard. The question worth bringing to your next team discussion: what problem are we currently accepting a good-enough answer on because the better answer is uncomfortable to stay in?
The Short Version (Pro): The individual can hold the line. The organization holds it only if the leader makes refusal visible, gives real voice to the people whose hesitation is worth trusting, and changes the review question from acceptable to earned. Premature closure is the default. Changing it is a deliberate act.
Try This Now (Pro): Before your next team review, pull one piece of AI-assisted output your team has already approved. Read it against the original problem it was meant to solve, not the brief it answered. Ask whether it earns the standard. Bring what you find into the meeting, not as criticism, as calibration.
References
1. MIT Research (January 2025). Hallucinating models use 34% more confident language than when generating correct outputs. 2. OpenAI (September 2025). Why Language Models Hallucinate — training objectives reward guessing over calibrated uncertainty. 3. AllAboutAI Comprehensive Study (2024). Global business losses attributed to AI hallucinations: $67.4 billion. 4. Karpowicz (2025). Mathematical proof that hallucination cannot be fully eliminated under current LLM architectures. 5. OpenAI o3 PersonQA benchmark (2025). o3 hallucinated 33% of the time, compared to 16% for predecessor o1.