For much of technological history, imagining something and being able to build it were activities separated by a considerable distance. A person could recognize a problem or conceive a company, but turning that intuition into an application, a service, or a tool required technical knowledge, capital, time, and almost always the collaboration of others; that difficulty never guaranteed that the result would be valuable, although it did limit who could attempt it. Artificial intelligence is shortening the distance because it already makes it possible to research, design, program, and analyze at a speed that a few years ago was reserved for specialized teams, even though its results still need supervision. The novelty does not consist in an idea having become sufficient, but in the fact that executing it is beginning to be considerably less exceptional.
Describing this change solely as a productivity story is incomplete. It is also a story about scarcity, because every technology that makes a capacity abundant displaces value toward what is still difficult: calculators reduced the usefulness of certain manual calculations without making mathematics unnecessary, while the internet expanded access to information without teaching us by itself how to understand it. AI covers a larger territory because it drafts, compares documents, proposes code, and offers explanations that appear reasoned. When a first convincing answer costs less and less, what becomes scarce is the capacity to decide whether it answers a question that matters, whether it deserves trust, and what consequences acting on it would have.
As a medical student in Ecuador, that difference does not feel abstract to me. Clinical training forces one to distinguish a plausible answer from a responsible decision, because a possible diagnosis does not have the same weight as a probable one, a general recommendation can be inadequate for a particular patient, and recognizing one's own limits is part of competence. The place from which we learn also modifies judgment: a model trained on data and priorities from other health systems can be useful, but it does not know by default the local availability of tests, medications, specialists, or follow-up. Added to this is an inequality of access that predates AI; the results of ENIGHUR 2024-2025 released by INEC in May 2026 placed internet access at 74.3 percent of the country's dwellings, with coverage of 78.3 percent in urban areas and 62.2 percent in rural ones. The idea that "everyone can build" describes a technological direction, not a uniformly distributed reality.
That caveat does not reduce the opportunity. In countries where capital and specialized talent are hard to assemble, cheapening the experiment allows a person without technical training to produce a prototype, a programmer to explore several solutions before committing to one, and small teams to attend to local needs that would never be a priority for a large foreign company. Technical ease, however, does not increase in the same proportion our capacity to recognize good problems; an observational study published by the NBER in 2026, based on data from more than 100,000 developers, found that the increase in programming activity associated with autonomous agents was sharply reduced when measuring finished projects and launches, while the larger number of applications did not produce an aggregate increase in use. It is initial evidence, not a general law, but it shows why executing more is not equivalent to creating more value. The old difficulty of building was an imperfect and often unjust filter; removing it does not select better ideas, it leaves the next filter exposed.
We can call that filter judgment, provided we do not use the word as an elegant substitute for "thinking well." Judgment gathers knowledge, experience, comparison, and a willingness to correct oneself; it makes it possible to observe how people actually behave instead of assuming how they should, to distinguish a persistent need from a striking anecdote, and to abandon a hypothesis when the evidence no longer supports it. It also does not appear by accumulating answers or by learning a collection of clever instructions for conversing with a model, because it depends on having seen enough cases and understanding the relevant mechanisms. This is the most tempting confusion of abundance: accessing an explanation seems equivalent to being in a position to evaluate it, although to detect an incorrect answer we need to know something, and to choose between two reasonable alternatives we need context before the moment of deciding arrives.
Medicine makes that limit visible with an uncomfortable clarity. A system can enumerate differential diagnoses, summarize a clinical history, or draft comprehensible recommendations and still omit a decisive datum, reproduce a bias, or induce excessive confidence in a well-written answer; the World Health Organization has warned about false or incomplete information and about automation bias, which appears when professionals or patients let pass errors they would have detected without a machine's suggestion. This does not justify keeping AI out of health, where it can help in education, documentation, research, and well-defined clinical tasks, but it does require preserving a distinction that holds for any profession: delegating a task does not automatically transfer responsibility for its consequences.
Education faces a less immediate version of the same problem. For a long time we evaluated learning through results because producing them was a reasonable approximation of the process that made them possible: an essay allowed one to observe a certain capacity for argument, and a working program suggested some understanding of the code. AI weakens that relationship by offering a polished result without requiring the decisions, doubts, and corrections that the task was meant to exercise, so that an impeccable text no longer demonstrates by itself that its author knows how to construct an argument, although it does not demonstrate the contrary either. It has lost part of its value as evidence. The educational difficulty then consists in specifying what capacity we wanted to observe and designing an evaluation that still makes it possible to see it.
A randomized trial published in 2025 in the journal PNAS, conducted with nearly a thousand high-school students in mathematics classes, showed that separation concretely: access to an open GPT-4 interface improved performance during practice, but those who used it obtained grades 17 percent lower than the control group when they later sat the exam without help. The harm largely disappeared with a tutor configured to offer hints designed by teachers instead of delivering answers, although it also produced no significant later improvement. A study in one subject and one context does not allow generalization to all of education, but it does show that the design of the assistance matters as much as the power of the model.
Banning these tools is an understandable reaction when the student delegates precisely the skill that was supposed to be practiced, but trying to preserve intact an education prior to AI would be a fragile strategy. There are efforts that are worth eliminating and others that seem inefficient only because we are looking solely at the result: trying to remember before searching reveals what we know, solving a problem badly helps locate the point at which the reasoning went astray, and writing a deficient first version forces one to discover that an intuition was not yet an argument. In my training, formulating a hypothesis before consulting the answer is not a nostalgic ceremony; it serves to order probabilities and make visible gaps that an immediate solution can hide. If we remove all friction, we run the risk of also saving the effort through which the capacity we later hope to delegate with judgment is built.
The alternative requires something more than allowing AI and asking for a footnote. An evaluation can require justifying decisions, comparing versions, explaining what evidence would change the conclusion, and pointing out where the tool is likely to be wrong; a brief oral defense or the analysis of a case with incomplete information usually reveals more understanding than an automatic text detector. It is also worth declaring what part of the work was delegated, verifying sources, and answering for the final version, not as procedures meant to punish technological assistance, but as habits of intellectual responsibility. UNESCO has proposed a people-centered adoption attentive to agency, inclusion, and pedagogical design. The correct proportion will vary according to the subject and the moment of training, because it does not make sense to offer the same help to someone who is learning a foundation as to someone who has already demonstrated mastery of it.
The reduction of barriers can also bring learning closer to reality. In many digital domains, where experimenting is cheap and reversible, a student learns programming while trying to create an application, understands economics when discovering that no one is willing to pay for their idea, or improves a design upon observing that users interpret an interface in ways no presentation had anticipated. This possibility is especially interesting in Ecuador and other Latin American countries, where students and small teams know nearby problems that global products barely register; theory remains necessary, but it meets real constraints sooner. Building for another person forces one to discover that something must work outside a demonstration, that context modifies the solution, and that an initially convincing idea may not survive contact with evidence.
Not all experiments, however, admit the same margin of error. Testing a tool to organize shifts, translate health information, or reduce administrative load allows iterations that would be unacceptable if the system were to guide clinical conduct; there, validation, data protection, professional supervision, and a strict definition of success are needed. The difference depends on the possible harm, the reversibility of the error, and who retains the obligation to answer, a caution proper to Medicine that also serves to evaluate decisions in finance, law, education, or public administration. Technological culture tends to celebrate speed as if it had value in itself, while high-risk professions remember that a fast decision is better only when it preserves safety and context.
When creating a technology company requires less capital and less initial technical knowledge, more experiments appear, but the number of relevant human problems does not grow at the same rate. Competition shifts toward trust, distribution, access to pertinent data, and situated knowledge; for a health project, for example, it is not enough that the model work in a demonstration if it ignores the workflows, the language, the regulation, or the constraints of the system where it intends to be used. Cheapening failure helps because it makes it possible to abandon a weak hypothesis sooner, although it introduces the inverse risk of confusing activity with progress: launching ten products does not produce ten times more learning if none began with a sufficiently good question. Speed acquires value when it shortens the time between a decision and the evidence that makes it possible to correct it.
Something similar will happen with the skill of using AI. Formulating precise instructions and designing workflows offers an advantage today, but interfaces will improve and many techniques that now seem sophisticated will end up incorporated into the products; deep knowledge of a field will remain hard to compress because it makes it possible to formulate less obvious questions, recognize exceptions, and understand why a solution that is correct in one context can be irresponsible in another. That is why I am not entirely convinced that the main task of education is to prepare students for "the jobs of the future," an expression that presupposes that we can anticipate occupations and tools that will change several times during a professional life. It seems more sensible to form people with foundations for learning, autonomy to work without permanent instructions, and honesty to recognize when a convincing answer exceeds their understanding.
AI can reduce one of the great barriers to innovating: the distance between an idea and its execution. It would be absurd to preserve that difficulty only because we learned to live with it, although eliminating it leaves in view responsibilities that could previously be confused with technical limitations. We will have to decide which problems deserve attention, which information is reliable, which efforts still form capacities, and which decisions should not be handed to a system incapable of answering for its effects. Education will no longer be able to content itself with teaching how to produce scarce answers, because answers are beginning to be abundant; it will have to occupy itself with the distance that still remains, the one that separates a convincing answer from a responsible judgment.
