Certo na Teoria, Inútil na Prática: IA, Dívida de Conhecimento e o Sujeito Oculto da Liderança Técnica

The current literary trend, dubbed "impeccable diagnosis," has gained popularity over the past two years. It begins with a correct distinction, presents a factual point, and concludes with a striking statement about a new role for engineers. This rhetoric, in line with Aristotle, discusses what should be done, but strangely, the subject of action rarely appears.

The most common example argues that small, disposable systems can be entirely entrusted to AI, while legacy, critical, and integrated systems require close supervision. Someone must monitor the generated code, understand its impact on architecture, validate security and scope. Leaders in technology must separate disproportionate resistance from legitimate risk signals.

The software engineer's role shifts from coding to deciding, for each system, the level of autonomy the code can have. While each proposition is true individually, the gap between a correct diagnosis and an organization capable of executing it is troubling. It echoes a Kantian text, where the hidden subject is strategy, a syntactic construction that allows making obligations without assigning them, describing work without budgeting it, and transferring responsibility without transferring the ability to execute it.

The question isn't whether someone needs to monitor, but who, with what prior knowledge, in what fraction of the day, and with what authority to say no? If the answer is the team, it's worth checking if that team knows the system's architecture, the decisions that led to its peculiar coupling between modules, the reason for the awkward distributed transaction no one dares to touch, and the meaning of the table with the wrong name that has existed since the 2019 migration and three integrations depend on it.

If the answer is no, and often it is, because the team joined later, inherited the code without inheriting the context, and received AI as a compensation for the lack of onboarding, the phrase "someone needs to monitor closely" doesn't describe a process. It describes hope. In 1793, Kant published an essay titled "On the Doctrine of Common Sense," which, while true in theory, is useless in practice.

He criticized the favorite excuse of pragmatists throughout history, the idea that theory is beautiful but reality is different. Kant argued that if a correct theory doesn't work in practice, the problem isn't in practice, but in the theory, which is incomplete. It's missing the middle term: the faculty of judgment (Urteilskraft), the ability to subsume the particular case under the general rule.

The cruel detail is that Kant believed this faculty couldn't be filled by more rules. If you add a rule to explain how to apply the first one, you'll need a third one to explain the second, and so on ad infinitum. The judgment is a natural talent that develops through concrete case practice. That's why, as Kant observes elsewhere, a doctor can know all the theory and still make the wrong diagnosis in front of a patient.

In 2026, the policy of supervising AI in critical systems is a correct theory. If it doesn't work in your company, the problem isn't that developers are lazy or resistant. It's that the theory is incomplete, lacking the specific knowledge of that system. This knowledge doesn't come from an alignment meeting, an outdated architecture document, or a prompt context.

Popper emphasized the importance of observation, suggesting that without a prior expectation organizing the observation, there is no observation. Hanson formalized this in 1958, stating that every observation is theory-laden. Applied to an AI-generated diff, it becomes an operation. Two people review the same thirty-line change.

One knows the system and sees that the change introduces a second source of truth for the order status, which the new join will pass through a growing table of thirty thousand rows per day, and that silent retry will generate duplicated charges for one of the three queue consumers. The second person sees thirty lines of syntactically correct, well-named code with passing tests.

The first revised, the second observed. Determining that the team needs to monitor the code being generated without first building the theory that makes observation possible is to give Popper's order and wait for science. For Aristotle, prudence, or practical wisdom, is not knowledge of universal principles. It's the ability to make sound decisions about the particular.