You have had this experience. A question arrives that you would once have sat with for twenty minutes. Instead you type it, read what comes back, find it reasonable, and move on. The whole thing takes ninety seconds and you are not aware of having decided anything.
Researchers at Wharton gave that moment a name this year: cognitive surrender: accepting a machine's output with minimal scrutiny, overriding both intuition and deliberate reasoning. Their finding is not that people are lazy. It is that the sequence matters. When the answer arrives before the thinking, the thinking mostly does not happen, and nothing in the experience signals that anything has been skipped.
This is an article about your own judgment rather than your organisation's, because the erosion happens one person at a time and the person it happens to is the last to know.
What the research actually found
The headline results are unambiguous. A study of several hundred adults found a strong association between frequent AI use and lower scores on standardised critical thinking measures, with the effect concentrated in younger and heavier users. Broader reviews this year describe the same mechanism from different angles: attention-capturing design promotes offloading, offloading reduces sustained analysis, and reduced analysis shows up as weaker autonomous judgment.
But the finding worth your attention is the one that complicates the headline. The research consistently splits by mode of use, not volume of use. Passive use, where you accept the output and move on, tracks with skill decay. Structured use, where you delegate a defined category of work and then direct the freed capacity at something the machine cannot do, tracks with the opposite: higher vigilance toward AI output and better transfer of learning. Same tool. Same hours. Opposite outcome.
The question is not how much AI you use. It is whether anything in the way you use it obliges you to reason against what it gives you.
Why judgment is the thing that decays
Judgment is not knowledge. It is not the set of facts you can state, and it does not sit in memory the way a phone number does. It is a practice: the accumulated result of having weighed incomplete evidence, committed, and found out. Every occasion you do it, the capacity is slightly reinforced. Every occasion you skip it, nothing is subtracted, which is exactly why the loss is invisible.
What automation removes is not your ability to judge. It removes the occasions on which you would have practised. Routine cognitive work was never valuable as output: the middling analysis, the first pass at a problem, the summary you would have had to write in order to discover you did not understand the thing. It was valuable as repetitions. Hand over the output and the repetitions go with it.
The uncomfortable part is that the work most worth delegating and the work that builds judgment are frequently the same work. Nobody defends the value of a first draft nobody reads. But writing it is how you find out where your understanding is thin, and reading a good one that arrived instantly is not.
Why you cannot feel it happening
Skill loss in a physical domain announces itself. You return to an instrument after a year and hear the difference immediately. Judgment gives you no such signal, because the faculty that would notice the decline is the one that has declined.
Worse, the experience of using AI well and using it badly is nearly identical from the inside. Both feel fluent. Both produce something plausible. The difference is whether you interrogated the output, and interrogation is effortful in a way that fluency actively discourages. This is why self-report is useless here and why the people most confident that they are using these tools critically are not reliably the ones who are.
The question is not how much AI you use. It is whether anything in the way you use it obliges you to reason against what it gives you.
The practice that holds it open
The structured use the research points to is not complicated, and it is not abstinence. It has one property: it forces a position before the answer arrives.
Form a view first, even a rough one, before you ask. Then the machine's output becomes something you compare against rather than something you receive, and any divergence is information: either you were wrong, which is worth knowing, or it is wrong, which is worth knowing sooner. Ask it to argue the opposite of what it just told you and notice whether the second case is as convincing as the first; if it is, neither was reasoning. Keep the work where being wrong has consequences on your own desk, because that is the work that was building the capacity in the first place.
And take seriously the category the machine cannot enter at all. It operates on what has been expressed. The judgment that matters most in your work concerns what has not been written down: what a hesitation in a meeting meant, why a plan that reads well will not survive this particular set of people, what your own experience is telling you that you have not yet articulated. That territory is not contested by AI. It is only abandoned, quietly, by people who have stopped going there.
Why this becomes an organisational problem
Individual erosion would be a private matter if it stayed individual. It does not, because everyone consults the same few models and receives the same shape of answer. The result is convergence that arrives without any of the friction that normally makes groupthink visible: no dominant voice, no suppressed dissent, just a room of people who independently arrived at the same reasonable position and take the agreement as confirmation.
That is the organisational form of this problem, and it has its own article. The individual form is the one you can act on today, and it is upstream of the other. An organisation's capacity to think well together is not separable from whether the people in it are still practising the thing on their own.
What is actually at stake
Nobody is going to stop using these tools, and the argument for stopping would be a bad one. Getting faster is real, and the speed is not an illusion.
But speed and judgment are separate accounts, and only one of them is being credited automatically. The reps that built your judgment were never scheduled. They were embedded in work that has now been automated. If you want to keep the capacity, you will have to put them back deliberately, because nothing in your day will do it for you and nothing will tell you when they have gone.