Exploring AI Bias With Teachable Machine

Purpose & objective

In this session, teachers will create an AI model using Teachable Machine (a free, open-source, online tool developed by Google). This activity is part of a lesson developed by Computer Science for All NYC which is appropriate for students in grades 6-12. Teachable Machine allows users to create different types of AI models, but for this activity, we'll use the image classifier to make an "emotion recognition model". Participants will train the model on a dataset with images of their own faces expressing different emotions (note that for data privacy reasons, these images are not saved by Google or used for commercial purposes). When participants test each other's models, they will immediately notice that their models won't work for everyone! At that point, we will interrogate the input datasets for the models. The purpose of this session is to demonstrate how bias in training data can lead to inequitable outcomes. At the end of the session, participants will be able to identify three key methods to improve training datasets: diversify the data, use a larger sample size, and norm the data by reducing confounding variables.

Outline

Opener/Do Now (5 minutes): Explore the AI tool Quick Draw
Demo (10 minutes): Make an AI model in Teachable Machine
Activity (20 minutes): Make a model in Teachable Machine, test other models
Discussion (10 minutes): Discuss why models failed, provide suggestions to improve the model
Video (5 minutes): Watch a video about emotion recognition AI
Discussion (10 minutes) Discuss the implications of emotion recognition AI in and outside the classroom

About half of the session will involve using the Teachable Machine tool, and the rest of the session will be divided between direct instruction and peer-to-peer interactions.

Supporting research

This workshop is inspired by other lessons in the Daily AI curriculum, which was developed by education researchers at MIT. (https://education.mit.edu/project/everyday-ai-for-youth-edai/#overview).

IEEE published a paper in 2020 on Teachable Machine's feasibility in the classroom (https://ieeexplore.ieee.org/document/9156030).

Our team ran an Equity + AI summer intensive for 50 teachers in 2023 which featured this activity, and exit survey data showed that teachers were excited to implement this lesson in their classrooms, and confident in their ability to do so.