Interactive and Safe AI for KS2 upwards

Learning how to use AI responsibly requires foundational knowledge about how AI works, an understanding of AI's shortcomings and an abundance of critical thinking skills.  Awareness is also needed of the ethical and environmental issues around AI's use and the role that humans have to play in AI's development, from selecting fair training data through to being accountable for their AI's decisions and outputs.  This is AI Literacy in a nutshell and we have learnt from our experiences of social media that children need to know about AI's risks long before they start using the technology unsupervised.    

Yet there are big challenges to introducing AI literacy in the classroom, particularly in primary school.  Children under 13 cannot legally use online generative AI services (although Google  recently announced that under 13's can use Gemini, if linked to a family account), so it is hard for younger children to be given agency to safely learn about GenAI.  Even in secondary schools there may be policies prohibiting access to AI sites or a nervousness amongst staff about supervising live use of AI, no matter how worthy the learning objectives are. 

As an informal educator delivering AI workshops in schools, I needed a way of running interactive activities that let KS2 students upwards explore AI, but kept them completely safe.  The constraints I set myself were: 

1. The activities must not use any online services 
2. The equipment must not have any internet connectivity and must be fully locked down 
3. The activities must not require any coding (coding is still a valuable skill but should not be a barrier to AI Literacy) 
4. All data captured by students, and any AI models trained using this data, must be verifiably deleted at the end of each workshop.  
5. All Generative AI outputs must be guaranteed appropriate for school use. 
6. The presenter must be able to remotely configure, start, stop and monitor all the activities so that no time is lost while students login or setup environments 

As a visitor to schools, I am luckily able to bring my own set of equipment for students with me, running bespoke software, but the approach I have adopted may be of interest to teachers who are introducing AI experiments to under 13s.   

The equipment I provide to  each team of students is a Raspberry Pi 4 with a USB camera and an Android tablet locked down with Fully Kiosk.  The Pi runs off a Power Bank to make it portable.  A Pi 5 would be more capable however it runs hotter so would require a cooling fan, making it potentially less durable than a Pi 4 in a metal case.  All the Pi's and tablets are connected to my presenter laptop via a portable Wi-Fi hotspot, which has no internet connectivity making the network completely isolated.  The laptop runs a bespoke Flask Python application to orchestrate the activities across the Pis and tablets. 

The first half of a workshop introduces Machine Learning, for which experiments with computer vision are ideal.  Students begin by building an image classifier to recognise the difference between human faces and toy characters.  I was keen that students select and capture their own training data given how many of the issues with AI arise from the choices that humans make at the training stage.  Google's MediaPipe Python libraries allow Raspberry Pis to train and run an Image Classifier locally, along with many other vision tasks.   Older students could code the Python themselves, however I provide a simple user interface running on the tablets with buttons to capture a photo from the Pi's camera, review the images, train the model and test it with live video.  Whilst similar image classification can be done online using Teachable Machine, I introduce a competitive element by plotting barcharts to visually compare the accuracy of each team's models when tested against a common set of images.  From these comparisons, teams are motivated to refine their training data and improve the performance.  Educators should always aim to show AI development as an iterative process, where the humans must decide when the training data provides acceptable performance, as opposed to the AI determining its own behaviours.  At the end of the session I show all the training images and models being deleted, as reassurance to staff and pupils. 

The second half of the workshop focuses on Generative AI and I was keen to give all students the same freedom and experience, regardless of age.  My solution was to create an AI image generator in which students have access to a limited set of prompt words.  Using Stable Diffusion running locally I pre-generated 2,000 different images representing every possible combinations of prompt words and then moderated them all to ensure they are safe for school use.  Multiple variations of each image were created, by adjusting the seed values in Stable Diffusion, so that the results seen by each team are all slightly different.  This setup allows even the youngest pupils to create fantastic yet flawed pictures without needing to be individually supervised and whilst still operating safely offline. 

I am currently upgrading this activity to display AI-generated videos instead of images.  Pre-generating large numbers of videos beings its own challenges in both time and cost.  If you have a high machine with plenty of VRAM, you can run AI video generators locally, although I have ended up deploying WanGP to a Hugging Face Space, which gives cost efficient access to a GPU on a PAYG basis.  Using this setup you can generate 160 small video clips in about 8 hours for the cost of a few pounds.  Pre-generating content may feel like cheating, but the important point with GenAI is to critically evaluate the outputs rather than spend time watching them being generated.   Additionally, I can also guarantee some examples of bias in the outputs when, for example, all teams generate images of a footballer and see that most of them are depicted as men. 

With this custom setup I aim to squeeze the greatest variety of interactive learning experiences about AI into each session.  For the teacher, it frees them up from worrying about equipment availability,  hardware compatibility, networking issues, data privacy or online safety.  Running AI locally in the classroom brings many advantages and is only limited by your hardware capabilities.  However the outputs of local Generative AI still need to be moderated to ensure they are safe for school use. 

I would be interested to hear if you are using AI locally in the classroom or please get in touch if you would like to know more about the workshops, or the technical setup behind them.