A few days ago I was facilitating an AI workshop at a Swiss university as part of a conference. The conference was mostly about how to use AI (LLMs like ChatGPT in particular) to teach kids in schools. On the way back home I was thinking about what it actually means if we start to use AI for this purpose now. Is this really a good idea? And I realized that to be able to answer these questions I will have to clarify my own understanding of learning in humans first.
This text is the result of my reflections.
My arguments will be mostly based on ideas and terminology from machine learning. One reason for this is simply that I’m much more familiar with them compared to the concepts originating from psychology. The other reason is more substantial: in machine learning we actually try to implement learning processes, often based on our current knowledge about learning in animals and humans. Therefore these ideas are actually field tested (and they seem to finally work pretty well in many scenarios). And trying to build something is often a good way to understand it:
„What I cannot build. I do not understand.” - Richard Feynman
So let’s start.
Machine learning algorithms can be divided into the following types (beware: very concise explanations here! You might want to consult some other sources first, if you are completely unfamiliar with this topic):
- Supervised learning: We give the machine for each input the correct output („labelled data“) for learning
- Unsupervised learning: The machine learns the structure in the (unlabelled) data without any further hints (inputs)
- Reinforcement learning: The machine learns to maximize a reward signal. This reward signal is supposed to guide the machine to achieving a certain goal (e.g.: optimizing profit by trading on the stock market).
All three learning types lead to the construction of a simplified model of the data. In the case of humans, the data is what we perceive: signals from our senses (eyes, ears, sense of touch etc.). We assume that this data is generated by some kind of truth: the world around us. Note: in reinforcement learning the body is considered a part of this world. The world therefore includes the interactions of our body with other parts of the world.
This simplified model of the world is then used to predict the future states of the world. These predictions can then be used to optimize our behavior to achieve our goals.
All three learning types also play an important role in human learning. Some examples:
- Supervised learning: teaching a kid how to speak a foreign language
- Unsupervised learning: figuring out what a spoon is by just looking at it and playing with it for a few seconds
- Reinforcement learning: learning that eating an orange is much more fun than eating a lemon
Now let’s have a closer look at their properties:
- Supervised learning: it's impossible for the learner to assess whether the constructed world model is accurate or not. If the labels from the teacher are good, the resulting model will be good, if they are of poor quality, the model will be poor. But the learner has no information about the quality of the labels. And, because the teacher is required to label the data, he also controls the selection of training data.
- Unsupervised learning: It is actually possible to measure how well the internal model fits the data (for the experts among my readers: e.g. SSE for k-means clustering). But it remains difficult to estimate the performance of a given model in real life scenarios.
- Reinforcement learning: the learner can use internal metrics to measure how fast he improves in maximizing the reward. But: it might be difficult or impossible for the learner to understand the nature of the reward signal (which is derived from unknown goals). Therefore, the learner might only be able to measure his improvements in achieving a goal which he does not understand. Also it is not possible to estimate the minimal (worst) and the maximal (best) possible rewards: we never know how well we are doing on an absolute scale.
We see that both supervised and reinforcement learning have fundamental limitations we should be aware of.
But there are more issues to consider:
- The data the learner perceives is distorted by the perception process in the brain (to an unknown extent, therefore the distortion could be large)
- The data the learner perceives is only a very small fraction of the data which could be perceived (like by an imagined omni-aware being. „Partial observability“ in reinforcement learning terminology).
- The data could be created by another intelligent system (human or machine). Examples would be a book written by a human or the output of an AI. The relation of such data to the „true world“ might be much weaker than expected.
Understanding all these limitations is essential if we want to understand learning in humans.
Again some examples:
- Supervised learning: what do we actually learn about history if we are learning it from a north Korean school book? Or in the example above: just imagine a teacher pretending to teach French but actually teaching Klingon instead. If contact with real French speakers is impossible, the kids would have no means to detect the deception.
- Unsupervised learning: the model (constructed by our observations) powering our „opinion“ is too small (i.e. simple) to be really useful and by not being aware of these limitations, our „understanding“ does more damage than good. An example could be an oversimplified understanding of other cultures („French people eat mostly frogs“).
- Reinforcement learning: we do all the time many things to get what we want. But where does this wanting actually come from? Can it maybe even be influenced by me? Or maybe by „us“ (i.e. by changing the way we live together)?
Our current schools still rely almost entirely on supervised learning: a teacher is giving the kids some data (in the form of texts or numbers) and is teaching them how to operate on this data by telling them which outputs (answers, results of calculations, essays etc.) are correct/good or false/poor respectively. And if schools teach kids „how to learn“ they almost always mean „how to optimize the efficiency of the supervised learning process“. The other learning types are rarely included, nor do schools teach much about the significant drawbacks of supervised learning.
These days many school teachers and school related government officials talk about the dangers of using LLMs in schools: a significant fraction (some say about 20%) of their output consists of hallucinated material. Yes, this is indeed a problem. But nobody talks about the other 80%: the stuff everybody agrees that it's true, but might be hilarious BS nonetheless (doctors of the 19th century: „women can’t have orgasms“).
Supervised learning is actually „copy learning“. LLMs, for instance, are trained purely using supervised learning [1]. They did not experience anything they talk about by themselves (think about cooking recipes written by ChatGPT: the AI has - of course - no tongue and therefore no idea what it's talking about). They are not trained on any kind of reality but only on a shared world model of humankind (in the form of all the publicly available text data of the planet).
And why is supervised learning so popular in schools today? Several reasons come to my mind:
- In the past, the purpose of school was to produce capable professionals and soldiers. They need to learn all the skills required to operate successfully in a given economic/political/military context. For this, any kind of questioning regarding the source of the learned material (i.e. the model copied into the kid's brains) is simply not required to make the kids achieve these goals later.
- Parents and teachers are scared to admit that they don’t know something. If a kid asks „why is the sky blue“, they would say something like „because air is slightly blue and if you look through a lot of it, it looks blue“. This is - of course - not even an attempt of an answer because it answers the question with the question (and is also wrong in other ways). It’s not a problem that most people don’t have a better answer to this question. The problem is, that we are so scared to admit this to kids. It might be that teachers fear a loss of authority. But is it really helpful if they present themselves as sources of universal truth? Or parents might think that kids don’t feel well protected if their parents don’t know everything. But does such a strategy really help to prepare them well for this world?
- Nobody is really interested in kids who are able to reflect on their learning process. The reason is that schools have become a battlefield of different groups in society who want to control the future by filtering and shaping the stuff kids learn (i.e. learn to simply believe) in school. Such control is only possible if the kid’s „learning skills“ are strictly limited to supervised learning.
- Existing dogmas determining what kids should believe. If we, for instance, insist that the „Book of Genesis“ is literally true, we also can’t accept modern biology. And - following something like a „reasoning chain reaction“ - we have to ultimately reject all science (including the scientific method for knowledge discovery).
- „Free thinking“ is considered dangerous: what if kids draw the „wrong“ conclusions (for instance: adopting an „extremist“ political ideology)? Our society declares many topics „off limits“ for free thinking. This leads to kids getting scolded if they ask certain questions.
The last point is particularly interesting: Isn’t it then better to limit their thinking to a set of „generally accepted harmless“ ideas? In this case they cannot make mistakes which could harm society.
The problem is again that in this case learning must be strictly „supervised only“. Modern schools often pretend to teach their kids „critical thinking“ these days. What they usually mean with this is „learning to find the flaws in other peoples arguments“. This should be (according to these schools) done by verifying claims by comparing them to „high quality“ sources of knowledge which are generally accepted to contain „true“ knowledge (like Wikipedia or „quality newspapers“).
Also many schools teach their kids to „verify“ the output of AI chatbots in this way (because of the hallucinations mentioned above).
This, of course, cannot work, as such platforms and media channels are plagued with the same fundamental problems of all other sources of knowledge (even if maybe to a lesser extent).
The only way we can truly verify knowledge is to compare it to our own experiences. This does not mean that we have to discover all our knowledge by ourselves or that we should not trust anybody. This would be, considering the vast amounts of human knowledge, not feasible. But it needs awareness and understanding regarding the different sources of knowledge. Like, for instance, how to use trust in other humans and machines wisely (trust/distrust is a result of repeated interactions - i.e. experiences - with other intelligent entities). How to assess the quality of the world model we are slowly building in our brain: when can we trust it (and make decisions based on it) and when rather not?
Should we therefore forbid the use of AI tools in schools? No, AIs are already part of the kids environment and therefore they must learn how to use them. And supervised learning (I really prefer the term „copy learning“) is not bad per se as long as the learner knows what he is doing, what the limitations and pitfalls of the process are. It's one of the most powerful abilities of humans that we are able to learn from other humans. This process is - in the first stage - always „copy learning“ and this is not necessarily a bad thing.
But if we don’t teach the kids to understand the learning process itself, AI becomes a very serious threat for society. While most kids learn sooner or later that humans (even teachers and scientists) are not always correct, this might not be true anymore in the case of intelligent machines. Kids might have far too much confidence in the extremely well presented outputs of a machine which was actually trained only on data created by humans (which are by no means less fallible than the kids).
Also, in the age of the internet information flow has become extremely difficult to control. Kids therefore always come - sooner or later - in contact with „alternative“ (in the sense of „not coming from the official source“) narratives about reality. If this happens, it's important that they are well prepared.
The alarming popularity of many conspiracy theories shows that today too many people are not well prepared: the day these people suddenly realize that the „official“ narrative - carefully constructed over many years of learning in schools, watching TV and reading the local newspaper - might be seriously flawed in some way, they immediately start to believe in another, seemingly „truer“ narrative. For too many of us the fear of not being in the possession of the full truth is so scary, that any gaps in their understanding of the world must be immediately filled, even if the „filling material“ is obviously of questionable quality.
Therefore, understanding learning also means to accept and embrace the state of „not knowing“.
Ultimately „not knowing“ is a beautiful thing, as it opens the space for dreams.
[1] Later training stages sometimes include reinforcement learning (for alignment, optimizing for user preferences etc.)
21.05.25:
I have, a few days after writing this text, decided to start a project related to the ideas presented here:
- know-know.org (an initiative to foster learning awareness)
Image: Shutterstock, artist: „MaksEvs“
Follow me on X to get informed about new content on this blog.
