An open project by Marco Iannacone that originated Proxima ZSP
Working draft, open to critiqueThe right to think for yourself
A manifesto for an artificial intelligence that helps you think, not one that thinks for you.
The fact, before anything else
One teenager in three, worldwide, already uses an artificial intelligence tool every week to do homework. In the vast majority of cases, that tool does one thing only: it answers. It writes the essay, solves the equation, summarizes the chapter. It does this well, in a few seconds, without ever asking the student to stop and think.
In 2025 a group of researchers measured what happens when the tool is taken away. Nearly a thousand high school students worked for weeks with an AI math tutor. Those who had used a version that supplied the solutions improved visibly during practice, but collapsed on the final unaided test, ending up performing worse than students who had never used the tool at all. Those who had used a version that offered only hints, never the solution, showed no such collapse. Same students, same subject, same underlying technology. The difference was entirely in the design.
A systematic review published recently, surveying thirty studies across thirteen countries, reaches the same conclusion by an independent route: artificial intelligence almost always improves immediate performance. It improves genuine learning far more rarely, and only when someone has designed the system so that it never offers the shortcut.
The principle we already know
This is not a new problem. Eighty years ago the psychologist Lev Vygotsky described the distance between what a child can do alone and what the child can do with an adult's guidance. He called that space the zone of proximal development: the exact place where learning happens, if someone guides it without doing the thinking for the child.
The principle is old; the threat is new. An adult helping a child gets tired, has limits, and normally stops before fully substituting for the child's own thinking. A language model never tires, and by its very design tends to do the opposite: solve the problem all the way through, in the fastest and most satisfying way for whoever is asking. Without a designed constraint, that tendency is not neutral. It systematically pushes toward replacing thought rather than supporting it.
Institutions are starting to notice. UNESCO has recommended since 2023 an age threshold below which independent use of these tools should be avoided. The EU's AI regulation classifies as high-risk any system that decides access to educational pathways, evaluates learning, or steers a student's course of study. Norway has chosen to ban generative use under thirteen and to supervise it up to sixteen. These are important steps. None of them, however, answers the simplest and most urgent question: how do we know whether a specific system helps reasoning or hinders it?
Today, that question has no public answer. At best, it has an answer held by the company that built the system: one that grades its own homework, by its own criteria, with no obligation to make them verifiable.
What we propose
An open framework, published under a free license, that anyone can read, verify, criticize, and apply to any system, including our own.
It is called L×M×C, and it measures three things that, taken together, distinguish a tutor that helps you think from one that thinks for you.
Depth of reasoning
Did the student merely copy a shortcut, or did they actually walk through the logical steps needed to reach the solution?
Metacognition
Is the student aware of how they are reasoning, does it recognize where it got stuck, can it say why a step is right or wrong?
Consolidation
Does what the student learned today, with the tool's help, still hold up tomorrow, when the tool is gone?
None of these three things is measured by whether the exercise was completed or the grade was high. They are measured by looking at how the student got there, and whether that path leaves a trace that lasts over time. This is exactly the distinction that 2025 research made visible: performance during use does not predict learning after use. Learning must be measured directly, not through its more convenient stand-in.
The framework is published open access, under a Creative Commons license, with a permanent digital identifier, ready to be read by anyone. It is not a product. It is a measuring instrument. And an instrument, to be one, must be able to measure whoever built it too.
What we ask
Of researchers
read it, put it to the test, find where it fails. A rigorous critique is worth more than a hundred compliments. If a piece of the framework doesn't hold up to your scrutiny, tell us, and let's improve it together. If you want to extend it to another domain, another age group, another language, the work is already as much yours as ours.
Of teachers
demand this measure from any tool that enters your classroom. Don't settle for the word of whoever is selling the product. Ask: how does this system know whether it's helping my students reason, or simply doing their thinking for them? If there is no verifiable answer, the question stays open; closing it is not your responsibility alone.
Of policymakers
a regulation that classifies a system as high-risk needs a measuring instrument to verify it, not just a list of documentation requirements. A public, replicable framework, tied to no vendor, can be that instrument. Don't wait for it to be offered by whoever has an interest in passing the exam.
Of whoever has read this far out of curiosity
share it. Ask the same question at your child's school, of the teacher, of the platform they use. Every time someone asks it, the question becomes a little harder to ignore.
Why now
In a few months, across Europe, AI systems used in schools will have to account for how they are built. The exact timeline for those obligations is still being defined, but the direction doesn't change: schools will need public criteria before it becomes mandatory to apply them, not after.
Meanwhile, every day, millions of teenagers open a tool that answers. The problem doesn't wait for the regulation to be ready. A student who delegates reasoning to a machine today doesn't get that reasoning back when the law, a year or two from now, asks them to demonstrate it.
We are not proposing to ban artificial intelligence for young people. We are proposing that no tool enter a classroom before someone has been able to publicly verify whether it teaches students to think or stops them from trying. It is a question within reach of anyone willing to ask it. It is time someone started demanding an answer.
The framework's code is open. The method is public. The data, once it exists, will be public too. We are not asking for trust. We are asking to be verified.