By Hildeberto Mendonca, on Saturday, March 28, 2026
Darius Williams is a teacher at Wheatley High School in Houston, Texas. He broke down in tears, describing a distressing moment during a class with high school seniors. The educator attempted a simple reading and writing activity with 17 and 18-year-old students. Despite reading two short paragraphs together, annotating the text as a group, modeling the exercise, and providing sentence stems where only four words needed to be filled in, the students were unable to complete the task. The teacher noted that while these students possess strong reasoning and verbal skills in informal settings, they struggle significantly with foundational literacy when asked to apply those ideas to written text. And that was only the second day of the school year.
The context where Mr. Williams is inserted has nothing to do with AI, but it helps to illustrate what can happens to a larger population when students use AI to complete their assignments, reducing their time reasoning about the problems. Actually, the transformation is already happening.
Dr. Jason Gibson is a college professor in Mississipi, who implemented a trap to uphold academic integrity. 32 out of his 35 students failed a midterm test for using AI to generate their answers without proofreading.
To catch unauthorized AI use, Dr. Gibson hid the word "Madagascar" in white text within the assignment prompt — making it invisible on the page, but readable to an AI when students copied and pasted the prompt into a tool. The topic was the Industrial Revolution and the appearance of "Madagascar" would make no sense, exposing when a student had generated and pasted their response without reading it.
If 32 out of 35 are not reading the prompt, just copying and pasting it, and not even reading the generated text, and they are doing it in a context where they are not supposed to do it, imagine people in a context where they are stimulated to use AI and the usage is actually a performance criteria? How concious about their ignorance these people will be? That's what happens in Software development, which is one of the activity most impacted by AI. It is where most productivity gains come from. But so far, despite all predictions, AI still doesn't give enough confidence to replace people, until the next generation of engineers who won't be able to even understand what AI is generating.
AI is a service, and as a service it can go offline. When it goes offline, the productivity sinks. A database is also a service, but it can run locally. It is very expensive to run an AI model locally. When we use a service like a database, there is no mind decay, but the more we use AI, the less we
There is a compromise to make. While AI increases productivity a lot, some of the gained time needs to be invested in the engineer, who has to understand and criticize the code instead of accepting it blindly. This investment will pay off in the occasions when the AI is out of service.
So yes, when we develop software, we will be way more productive than before, but less productive than someone who only uses vibe coding and doesn't take the time to understand and criticize what is generated.
For repetitive tasks, don't ask AI to do it. Ask AI to generate code that does the repetitive task, and run that code when you need the task done. I realized that when I asked AI to restore a database from a backup and it did it well, but I knew I would do it again in the future, so I asked AI to generate a script that reproduces the same steps, and now I can run that script whenever I need to restore a database from a backup, without using tokens.