I spent years in rooms where the stakes of misreading a signal were measured in lives.
You learn, in those rooms, to hear the difference between the practiced calm of someone running a script and the deliberate calm of someone who has nothing to hide. They sound similar on the surface. They feel completely different if you are paying attention to all four channels — not just the words.
The practiced calm is smooth. The vocabulary is precise. The answer arrives too quickly for the question. Nothing lands wrong. Everything flows.
The deliberate calm has friction in it. Small pauses. The visible weight of choosing the right word. The face that is not quite keeping up with the voice. The timing that tells you the person is actually processing rather than retrieving.
That friction was always a signal. It meant someone was actually there.
What the Models Removed
The current generation of AI language models are the most fluent text producers in history. Not the most accurate. Not the most truthful. The most fluent.
Fluency, in the technical sense, means the output sounds like confident, natural language — grammatically coherent, contextually appropriate, tonally consistent. A language model can produce fluent text about topics it has incomplete or incorrect information about. It can produce fluent text that contradicts fluent text it produced ten minutes earlier. The fluency is a property of the output layer. It tells you nothing about the accuracy of what is beneath it.
This matters because human beings use fluency as a trust heuristic.
We have always done this. The person who speaks clearly and confidently is processed — automatically, below the language layer — as someone who knows what they are talking about. This heuristic was approximately useful in a world where fluency required practice and practice required genuine knowledge. It is catastrophically unreliable in a world where fluency can be generated in milliseconds from whatever pattern best fits the prompt.
The Specific Risk for Leaders
If you are in a position of authority — organizational, pastoral, military, parental — you are surrounded by people who are increasingly using AI to produce the communications you receive. The emails, the reports, the proposals, the explanations.
Most of those documents will be fluent. Many of them will be accurate. Some of them will contain errors, omissions, or framings that the person sending them did not notice because they approved the output without running a synthesis layer of their own.
The old signal of fluency — this person knows what they are talking about — no longer reliably tracks to the person. It tracks to the model they used.
The implication is not that you should trust fluent communication less. It is that you need different signals now.
Ask for the reasoning, not just the conclusion. The model can provide conclusions fluently. The reasoning will reveal whether there is a human mind behind it or a pattern engine. Ask where the information came from. Ask what was considered and discarded. Ask the question that requires something the model cannot fake: the felt experience of being wrong and having to revise.
What Still Carries Signal
Friction still carries signal.
The pause before the answer. The willingness to say I do not know without immediately filling the silence with plausible-sounding content. The moment when someone's face and their words are slightly out of sync because they are telling you something that costs them to say. The update — the genuine revision of a stated position because new information actually landed.
These are things a language model can simulate. They are things a language model cannot produce.
The distinction is not always visible on the surface. It requires the same skill that distinguished the Blue Angels debrief from a performance: the ability to hear what is underneath the words, not just the words.
That skill is not obsolete. In a world saturated with AI fluency, it is more necessary than it has ever been.
Train it accordingly.