Not a Thinking Being, but Merely a Machine

Image: Students playing Hexapawn—a simulation of how computers appear to learn.

If the system is based solely on statistics, it cannot know the truth. It can only replicate patterns.
A machine has no morality.

We live in an age where machines speak, and many people prefer to rely on the output of AI systems rather than on the advice of other people when dealing with everyday problems. The current debate about artificial intelligence in schools often swings between blind euphoria and deep skepticism. But while we debate whether ChatGPT does homework, we often overlook one of the core issues: anthropomorphism. We tend to attribute human abilities such as «understanding», «wanting», or «thinking» to AI. This is dangerous. When children and adolescents perceive a machine as a counterpart that is «smarter» than they are, not only does their own sense of autonomy diminish, but the question also arises: «Why should we even bother learning anymore if AI understands everything better and can express it better?» This makes it all the more important to demystify the phenomenon of AI. The goal must be not to consume AI as a magical black box, but to see through it as a mechanical-mathematical construct.

Three levels of AI in education

When it comes to teaching AI literacy to students, we consider three levels to be important: how AI works, how it impacts society, and how it can be used effectively. The third level—the instrumental level—is heavily overemphasized in the current political debate on AI in education, while the other two levels are neglected. The paper by the Ständige Wissenschaftliche Kommission der Kultusministerkonferenz (Standing Scientific Commission of the Conference of Ministers of Education) on Large Language Models (LLMs) is an example of this. One aspect of the paper worth supporting is the call to prioritize the development of basic reading and writing skills before LLMs are used. However, the recommendation is limited to «learning with LLMs», rather than also incorporating «learning about LLMs» beforehand or alongside it.

In our view, this is a shortcoming, because levels one and two are crucial for responsible use of AI. We need to understand the long-term effects of AI use on learners in order to make informed decisions about whether or not to use it. The Massachusetts Institute of Technology’s study Your Brain on ChatGPT scientifically substantiates the adverse effects of using large language models—effects that many educators had already observed: cognitive abilities and memory are at risk. Other studies show negative effects on critical thinking. And that is only the visible part of the negative effects in front of the screen. It is equally important to look behind the screen. Examples include ecological consequences such as high energy consumption and social aspects such as the exploitation of data workers in the Global South during the training phase of AI models.

Learning what AI is (not)

But back to the first level. Paradoxically, the best way to demystify AI is to start with the end devices. The Computer Science Unplugged approach—that is, computer science without electricity—makes it possible to experience the logic behind the technology in a hands-on and social way. By slowing down the processes in the real world and taking a playful approach, significantly more students can engage with the subject. Understanding AI is also—and especially—important for those who consider themselves bad at math and drop computer science at the first opportunity. When AI principles are then reenacted using matchboxes, marbles, or slips of paper, they realize: No one is thinking here. Here, calculations are made, sorted, and weighted. It is pure mechanics, even if it happens at breakneck speed.

When matchboxes «learn»

A classic example that demonstrates the principle of so-called Feedback Learning is the game Hexapawn, which was developed as early as 1962 by the American science journalist Martin Gardner. In this game, a human player and a «matchbox computer» compete against each other on a 3x3 board. The computer consists of a row of matchboxes. Each box represents a possible game situation. Inside the boxes are colored beads that represent different moves. The human makes the first move. After that, the person operating the computer randomly selects a bead from the corresponding box for the current situation. The color of the bead determines its move. If the computer wins, it is rewarded. That is, the bead remains in the box. If it loses, it is «punished». The bead from the last move is thus removed. In this way, it can no longer make that wrong move that led to the loss.

After a few games, the board game computer becomes unbeatable. The students realize that, for an AI, «learning» simply means adjusting probabilities based on past successes or failures. There is no understanding of the rules of the game, no strategic genius—just statistical selection.

A network you can touch

While Hexapawn makes feedback learning accessible, the logical network—misleadingly dubbed Brain in a Bag by its developers—aims to foster an understanding of logical connections within a network of signal transmission.

In this game, students take on the role of individual processing units (PUs). They are connected to one another by strings. Each PU has a simple task: The PUs at the input level receive visual signals, which they compare to a «correct» visual signal. If the input and the reference match, they send a «yes» signal forward to the next level of PUs. These receive a different type of signal: «yes» signals in the form of cups or empty toilet paper rolls, which are sent from one PU to the next on strings. Only when the total number of signals exceeds a certain threshold is a «yes» signal sent to the next PU.

Various playing cards are held up, and the grid reacts accordingly. The rest of the students, however, don’t know the instructions and have to figure out how the grid works.

This helps students understand that AI does not store knowledge like an encyclopedia. Instead, it consists of billions of weighted connections. «Intelligence» here is nothing more than the correct transmission of signals through a system of thresholds.

Patterns instead of truth

Anyone who understands that a neural network merely calculates probabilities will take a more critical view of its answers. If the system is based solely on statistics, it cannot know the truth. It can only match patterns against probabilities. A machine has no morality. Responsibility for the outcome always lies with the person who selects the data and programs the machine. While AI is unbeatable at pattern recognition, it lacks the capacity for empathy, ethical judgment, and genuine creative impulse arising from nothing.

More than just operational proficiency

Waldorf education is about guiding students on their journey toward becoming responsible citizens and engaged members of society. This includes digital literacy, which goes far beyond mere technical proficiency. By exploring AI in the classroom in a hands-on, non-digital way, we can help students overcome their fear of the unknown and their awe of the supposed overwhelming power of algorithms.

We provide them with learning experiences that help them understand that behind the shiny surface of apps and chatbots is nothing but mathematics and logic. In doing so, we aim to empower them to use technology for what it is: a tool. Nothing more, but also nothing less. Discernment, empathy, and morality are human qualities.

As part of an AI-themed day, a twelfth-grade class explored the game Hexapawn and logical networks. In the evaluation form, under the heading «What I learned today», one student wrote: «… that AI isn’t a thinking being, but really just a machine.»