I was about to memorise a wrong formula, and my AI had taught it to me. It was the formula for invested capital, explained beautifully: clean example, clear steps, the confident tone of someone who knows. Before learning it I compared it with the course slide, more out of scruple than suspicion. The course version had one more piece. Since that day the scruple is a rule, for every definition.
I have used AI at work for three years and I used it to prepare the exams of my bachelor’s degree, completed in two years instead of three, while working. It never studied for me, as I have written before, and not as a matter of principle: because it does not work. Here I tell how I turned it into a tutor that quizzes me, and the mistakes it made on the way.
Why doesn’t rereading work?
In secondary school my method was read, underline, reread. At 41 it took me a month to realise it no longer worked. Rereading gives a feeling of mastery that vanishes at the exam, and that is not my impression. Roediger and Karpicke showed in 2006 that people who retrieve information through tests remember it better over time than people who reread it, even though rereading feels more effective while you study. Dunlosky and colleagues reviewed ten study techniques in 2013: retrieval practice and spaced repetition high utility, rereading and highlighting low. You learn by retrieving information, not by looking at it again. And to be quizzed every day, at any hour, I have found nothing more convenient than an AI.
How do you turn AI into a tutor?
The first step is telling it who it is. In every study project my fixed instructions said more or less this: you are my tutor for this exam, your job is not to study for me but to question me and correct me. The role is worth little without the rules, written one at a time and almost always after a mistake. Every explanation starts from the materials I uploaded, with file and page, and separates “from your materials” from “general knowledge”, because at the exam the two are not worth the same. If a piece of information is not in the materials it must tell me, not reconstruct it. Calculations it checks twice before giving me the solution. And when I get something wrong it does not give me the answer: it asks me a question that leads me to find the mistake myself, the most annoying and most useful thing a tutor can do.
With these rules every chapter became a fixed round. The AI explains one block anchored to the slides and right after, inside the lesson, asks me three to five quick questions. If I pass, a battery of questions with realistic traps follows, then a timed exam simulation, in the real format, several chapters together. Every mistake is classified and new questions come only on that type of mistake. You go back to the explanation only when the mistake is one of knowledge. All the others are cured by practising, not by rereading.
The tutor’s instructions and the five prompts of the round, with the names of the six moves (explain, checkpoint, drill, mock, diagnosis, re-drill), are written up in twelve pages, the Quaderno 3in2, which you can download here (in Italian). They work with any AI that accepts persistent instructions and your files.
What mistakes does AI make when you study?
Four in my case, all in the log. The first is the formula I opened with. The second is an answer key that, in a simulation, marked the wrong answer as right. The third is a “plausible” definition that differed from the course’s, and here the fault is more mine than its: I had not checked it against the source. The fourth is a mock exam with the correct answers almost all on the same letter, the fastest way to a good mark in a simulation and a bad one at the exam. More generally, the AI gets things wrong or “hallucinates” depending on the model you use, the context window and the skills you have installed, but that is a subject for a whole course…
From there came three fixed rules, which I call “don’t trust the AI”. Every exam-grade statement must have a cited source, technical definitions are checked on the slide before memorising them, and every now and then I ask it to audit itself: reread everything you have explained to me, compare it with the materials, tell me where you said something that is not there, where you used a term different from the course’s, where you simplified a formula.
Catching the mistakes of a tutor who speaks with confidence became, without my noticing, an exam skill. It forces you to know the source better than it does, and in the end you study twice as much.
What mistakes do you make?
The part that surprised me most is about me, not the AI. I kept a log of the mistakes made in the simulations, with a code: K if I did not know the concept, E if I knew it but got the calculation, sign or formula wrong, L if I misread the question (the “which one is NOT” questions are a massacre), D if I fell into a trap built on purpose, T if time ran out and I guessed, A if the AI had got it wrong.
On a multiple-choice and calculation exam, at an advanced stage of preparation, most of my mistakes were not type K. They were E and T, execution under pressure: I knew the subject and got it wrong anyway, for an inverted sign or one minute too few. Studying more does not help someone who fails like that. Simulating more does, in real conditions: same format, same time, difficulty above the real one, phone in another room.
In companies it is the same. A business plan rarely blows up for lack of knowledge and almost always for execution under pressure, when the numbers come in late and the decision has to be taken anyway.
Where is AI not enough?
Memory. It is the part nobody expects from someone who studied with AI. I could not memorise the graphs of microeconomics and macroeconomics by looking at them, neither in the handout nor in a chat. I redrew them by hand, on real sheets of paper (I am a nostalgic), in different colours, several times, until they came out on their own. Same for a particular formula, for the keywords of a law topic, for quick, ugly mind maps. Everything I wanted to sink in went through my hand at least once. The AI was for understanding and for being questioned, the pen for remembering, and neither does the other’s job.
Does it apply outside university?
Yes, and that is why I am writing it here and not on a student forum. If at work the AI prepares an analysis and nobody goes back to the source, the incomplete formula ends up in a board deck, with the same confident tone it reached me with, and there it costs more than an exam. The rules are the tutor’s: cite the source, separate data from assumption, check the calculations twice, and every now and then ask the AI to check itself.
I have already told the story of two numbers signed without going back to the source, in the first piece of this series: a mortgage and an exit, and I paid for both. With AI I learned in two years what in twenty I had not learned from people: a confident tone says nothing about the quality of the number. The nice part is that once you catch the habit of checking, you take it with you into meetings, financial statements and plans. And there the AI, strict as a tutor, is an exceptional colleague.
Sources: Roediger and Karpicke, “Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention”, Psychological Science, 2006; Dunlosky, Rawson, Marsh, Nathan, Willingham, “Improving Students’ Learning With Effective Learning Techniques”, Psychological Science in the Public Interest, 2013. The AI’s mistakes and mine come from my personal log; no university material is quoted.
