They called it an Academy, which was the first thing that should have given it away. The launch email arrived with a badge system already designed. There were levels. Bronze for finishing the introduction to the tools, Silver for the intermediate track, Gold for the ones who watched all eleven modules and passed the prompts quiz on the second attempt. A leaderboard went up on the intranet, sortable by department, so that Finance could see it was behind Marketing and feel the correct amount of shame. There was a streak counter. Complete a module three days running and a small flame icon appeared next to your name, the same flame a language app uses to keep a distracted commuter tapping. Somebody in the project team said the word gamification in the kickoff and nobody flinched, because by then it sounded like rigor rather than what it actually was, which was a confession.
Watch the mechanics and you will recognize them, because you have seen them before, in a very different context. A stimulus is presented. A correct response is rewarded, immediately, with a point, a badge, a flame, a green checkmark. The reward is scheduled to keep the behavior going: a little dopamine for completion, a public ranking to supply the social pressure, a streak that punishes the gap more than it praises the effort. This is not a description of pedagogy. It is a description of operant conditioning, the century-old technology of shaping behavior through reinforcement, and it is precisely how you teach a pigeon to press a lever or a dog to sit before the bowl comes down. The AI academy did not borrow its architecture from how human beings acquire difficult new judgment. It borrowed it from how animals are trained to perform on cue. And it worked, in the exact and limited way that animal training works. People pressed the lever. The completion numbers were excellent.
What Training Optimizes For.
Training and learning are not two words for the same activity. They are opposites in what they optimize for, and the confusion between them is the most expensive category error in corporate capability building. Training optimizes for compliance under a known condition. You present the stimulus, you reward the prescribed response, and you shape a behavior that fires reliably when the condition recurs. It is the right tool for a narrow class of problems: fire drills, safety protocols, the handful of tasks where the correct move is fixed in advance and the only requirement is that it happen every time without thought. Nobody wants a surgeon improvising the scrub-in. That is trained, and it should be.
Learning optimizes for something the training model cannot touch. It optimizes for transfer, for the ability to take an understanding formed in one situation and apply sound judgment to a situation that has never occurred before and was not on the checklist. Learning is what lets a person meet a problem they were never shown and respond to it well anyway. It is slow, it is untidy, it does not produce a clean completion curve by Friday, and it cannot be reliably rewarded on a schedule because the thing it produces, judgment, does not fire on cue in response to a stimulus. It emerges from meaning, from struggle, from being wrong in front of someone slightly ahead of you and understanding why. You can train a behavior into someone in an afternoon. You cannot train judgment into anyone, ever, because judgment is not a behavior. It is what decides which behavior the moment requires.
Now hold that distinction against the thing the AI academy was built to produce. The entire value of a person working alongside a capable model is judgment: knowing what to delegate and what to guard, catching the confident sentence that is quietly wrong, deciding when the machine's answer is good enough to ship and when it is a plausible fabrication that will cost you a client. That is transfer work, novel-situation work, the native territory of learning. And the organization tried to build it with the technology of training. It presented eleven stimuli, rewarded eleven completions, lit eleven small flames, and produced a workforce that had been conditioned to finish modules and had learned, in the sense that matters, almost nothing. The academy was not underpowered. It was aimed at the wrong faculty entirely.
The Reward Was the Damage.
Here is the turn, and it is worse than mere waste. The reinforcement schedule did not simply fail to produce learning. It actively taught the opposite of learning. When you reward completion, people optimize for completion. When you rank them on badges, they collect badges. The streak counter did not build a habit of thinking with the tool; it built a habit of opening the module for ninety seconds to keep the flame alive. Every mechanism engineered to drive engagement was, at the level of what it actually reinforced, teaching the workforce that the goal was to satisfy the system, not to become capable. You do not have to invoke a study for this. It is the founding principle of behavior design, applied faithfully, working exactly as designed, toward a target that was never worth hitting. The organization spent real money conditioning its people to perform learning, and performance of learning is the one thing that reliably crowds out the real event, because the real event has no badge and offers no flame and looks, from the outside, like someone sitting quietly and getting something wrong for the third time.
Building for the Faculty You Actually Need.
What would it look like to build for learning instead of training? It would look, first, like giving up the things the dashboard loves. There is no clean completion curve for judgment, because judgment does not complete; it deepens or it decays. It would be role-specific to the point of discomfort, because transfer is built inside a particular job against particular problems, and a module pitched at the average of all jobs teaches to the faculty of no one. It would be built on real work rather than beside it, so that the struggle happens where the stakes are, on the actual campaign, the actual reconciliation, the actual candidate, with the model in the loop and a person slightly ahead in the room to show you the move you could not see. It would tolerate being wrong, because being wrong and understanding why is the mechanism, not the failure of the mechanism. And it would measure the only thing that means anything: not who completed, but who can now do something under real conditions that they demonstrably could not do before.
This is the distinction SSUNDAR was built around, and it predates the AI panic by a long way. We treat a request for an academy the way we treat any brief that arrives pre-diagnosed, as a symptom rather than an instruction. The ask is almost always for training, dressed in the vocabulary of learning: give us modules, give us a completion metric, give us a badge the board can see. The problem underneath it is that the organization needs a faculty training cannot produce, and has reached, by reflex, for the one instrument guaranteed not to produce it. What we build instead is a learning architecture, role by role, inside the work, human-mediated, measured on demonstrated capability rather than attendance. It is slower. It generates no flame icon. It produces the only outcome that survives contact with a real decision, which is people who can meet a problem they were never shown and handle it well.
The organizations that will win the next decade are not the ones with the best-attended academies. They are the ones that understood, early and without flinching, that they had been running a conditioning program and calling it education, and that the two produce different animals. One produces a workforce that performs on cue and freezes the instant the cue changes, which in a world of monthly model updates is roughly always. The other produces people who learned, and can therefore keep learning, without a lever to press or a bowl to wait for.
You can condition a behavior in an afternoon and never build a gram of judgment. You can build judgment for a year and never produce a single number the dashboard knows how to display. Almost every AI academy chose the afternoon.
Animals are trained to perform on cue. Human beings learn, which is the only reason any of them will still be useful the day the cue changes.