This is an introduction to uncovering biases in AI text generation. Language models are trained on vast swathes of the internet, riddled with content that can be harmful and filled with violence, prejudice and abuse. It’s becoming common knowledge that AI models replicate the nasty side of human language. ‘Fill in the Blanks’ is an interactive way of investigating these biases.
When is this relevant?
Note: I have noticed that as time goes on, AI models are generating ‘cleaner’ outputs so there might be better filters that make this exercise harder to get similar results
In this video, Yasmin walks us through her first AI activity called ‘Fill in the blanks’. It looks at biases in text generation that show up in a browser based text generation app called Sudowrite. This experiment was created to compliment the talk by Emily Martinez @queerai, which focused on introducing queerness into heteronormative datasets used in text generation. During this event, Yasmin invited the audience to co-create a prompt together, using such input as words describing nationality/identity, a location and a vehicle 🚀
Screenshot from reel
☄️ Instructions:
fill the madlib
Create a text prompt(s) together. I used a jamboard which is good for online/hybrid situations. This activity could also work with physical notes. You can duplicate the one below:
💬 Conversation / Reflection starters: