Artificial intelligence, neural networks, and deep learning are topics surprisingly accessible to children. You don't have to cover the math and even if you do, it's not that complicated.
The best way to explain what's artificial intelligence is to provide a lot of practical examples that children can try on their own: Google Translate, ChatGPT, Alexa-like assistants, Socratic, DALL-E, voice-to-text on any phone, Google Lens, and little practical tools like background removal or super-resolution.
Once there are some examples, you get the opportunity to compare narrow and general artificial intelligence (which already exists). You can explain that artificial intelligence that learns from examples overcomes limitations of simple if-then-else programs.
You can then proceed with some hands-on exercises specifically designed for children:
- Machine Learning for Kids – The website might be a bit unwieldy and confusing, but don't let that discourage you. It provides an online dataset editor and generates classifier model, which can then be used as a block in the popular Scratch environment. There are many worksheets you can use in classroom or at home.
Explaining how artificial intelligence works is a bit more challenging, but it's not really that hard. All modern AIs are neural networks and single neuron is just weighted sum of evidence compared to threshold. My favorite example is a neuron that determines whether patient has COVID. All inputs represent evidence as logical zeroes and ones. You just have to choose weights. For example, positive PCR test would have a very high weight. Less reliable antigen test would have lower weight. General symptoms like fever and headache would have even lower weight. Some weights could be negative, for example positive flu test or a school exam scheduled for tomorrow :)
You can then explain that the point of training is to determine the weights automatically from lots of examples, which come from experts (in this case doctors). You can then introduce random and gradient descent by drawing some 1D and 2D landscapes and simulating both algorithms on them.
Network architecture can be quite intimidating, but there's no need to cover it in much detail. The trick is that neural networks will function correctly with some basic level of performance regardless of how you structure them. You can start with ordinary multilayer perceptron and get decent results. Network architectures are only relevant once you want to create the best model in the world for given task. You can explain some of the most basic building blocks of neural networks (vectors, matrices, activation functions) and encourage experimentation with architectures. Vectors will be easier to understand if you explain how images and classes are mapped to them (and perhaps also sounds and words).
Finally, you can mention that there are now thousands of tools and datasets, which make it much easier to train new models. Children will be relieved when they hear you can use these tools to create a decent narrow AI in a few lines of Python code.