1Review the content without delegating the review to the AI

The texts generated by artificial intelligence (AI) sound good, the explanations are convincing and the references look real. However, they are sometimes false. The risk does not lie in the obvious error, which anyone detects, but in the plausible error: the date that is almost right, the quotation that almost existed, the figure that almost matches. Martín Núñez Calleja, from the National Centre for Curriculum Development in Non-Proprietary Systems (CEDEC), sums it up clearly in his talk «Crear REA con eXeLearning en tiempos de IA» (Creating OER with eXeLearning in times of AI), devoted to open educational resources (OER): «In the classroom, an error that nobody detects is not an error: it is mistaken learning».

In an interactive resource the problem multiplies, because the error is not read once but repeated with every student who uses it. A simulator with a misapplied formula, a quiz that accepts a wrong answer as correct or an explanation that reverses a cause-and-effect relationship teach that error as many times as the material is opened.

It is advisable to review the whole material before publishing it, and not only the part that was expressly requested. The check covers four aspects:

• The content. The concepts, data, dates, units and formulas that appear in the explanations.
• The activities. The questions, the answer options and, above all, which ones are accepted as correct. A badly designed distractor can be as harmful as a wrong answer.
• The feedback. What the material replies when students get it wrong, which can only be seen by trying incorrect answers.
• Behaviour in edge cases. What happens when an answer is left blank, a negative number is typed, a huge value is entered or a button is pressed twice in a row.

The way to do this is to go through the material as students would, also with wrong answers, and not just look at the initial screen. Many faults only appear on reaching the end of the activity or on repeating it.

With AI it is easy to modify a resource that has already been reviewed, and that ease carries a risk. A change that seems small, such as adding a question or changing the design, can break another part that already worked, even if it apparently has nothing to do with what was requested. These faults are known as regressions, and the initial review does not detect them, since it was carried out on the previous version.

To detect them, the review becomes a checklist that is repeated after each significant change. The list includes the main walkthroughs and the edge cases that have already been checked, and the AI itself can draft it from what has been tested. For a quiz, for example, it could be the following:

• The quiz can be completed from beginning to end.
• A wrong answer shows the intended feedback.
• A blank answer does not cause errors.
• On restarting, the activity returns to its initial state.
• It works with the keyboard only.

When the project allows it, the AI can turn those checks into automated tests and run them after each change. Coding agents do this unaided. These tests only confirm that the material still works as before, and the correctness of what it teaches still depends on the review described above. The AI can take care of repeating the checks, although deciding which behaviours must be kept is up to the person who knows the material.

AI speeds up production, but pedagogical responsibility still lies with the person who publishes the material. This is also stated in the «Guía sobre el uso de la inteligencia artificial en el ámbito educativo» (Guide on the use of artificial intelligence in education) of Spain’s National Institute of Educational Technologies and Teacher Training (INTEF), in its version 2.0 of September 2026, which includes human oversight and responsibility among its ethical principles: teachers must keep control over the use of AI, and educational decisions cannot depend on automated systems.

This has a practical consequence for the rest of the guide. The AI can take care of almost all the recommendations, from the licence to accessibility, but not this one. It can point out what should be checked, and even warn about what it is unsure of, but it cannot certify that what the material teaches is correct. That is why, in the VCER evaluation, the AI scores only the errors it detects and points out what the person must review.

The review does not end on publication. The people who use the material find faults that its author has not seen, because they test it with other students, on other devices and with other questions. It is therefore advisable to make it known in a teaching community and to pay attention to the comments of the people who try it. In the Telegram group Vibe Coding Educativo, the «¡Comparte tu App!» (Share your app) section is meant for presenting published programs, and the comments they receive help to correct and improve them.

Accurate data are the minimum, not the goal. A resource can contain no errors at all and still be poor from a teaching point of view. The article «Mantener la “A” de abierto en los REA en tiempos de IA» (Keeping the “O” of open in OER in times of AI) points out three aspects that no code replaces.

The first is the clarity of the learning objective, so that students know at all times what they are expected to learn and why. The second is formative feedback, which does not consist only of saying whether an answer is correct, but of explaining why, offering a hint and inviting another attempt. The third is methodological coherence, that is, each activity responding to a teaching decision and not to what the AI proposed by default.

The question worth asking when reviewing is not whether the material has turned out showy, but whether it helps students learn better. That judgement belongs to the person who knows the subject and to the students, and it is their contribution to educational vibe coding.