AI Biases
Tue, Apr 14
One of the earliest critiques of AI that surfaced in the deep learning era was that modern machine learning pipelines tended to be biased when processing data . This has led to scrutiny of the training data these pipelines rely on, including its sourcing and curation. Human labor remains an important part of the process, and comes with its own ethical dilemmas related to bias and exploitation.
We have three main in-class learning goals. By the end of lecture today you will:
- Understand how bias affects machine learning-based AI systems deployed in commercial and government contexts.
- Reexamine the standard machine learning training regime to identify where human labor comes into play, and why that human labor can be both a source of bias and an exploitative practice.
- Ask whether bias is as large a problem as it was a decade ago.
The slides for today's lecture.
Read This:
Joy Buolamwini and Timnit Gebru published a landmark paper in 2018 highlighting accuracy disparities in commercial gender classification, which helped launch the entire area of the study of bias in AI.
Journalist Karen Hao takes us on a tour of where labels for supervised machine learning come from.
Do This:
Writing Reflection Extra Credit Assignment
See the instructions posted on the assignment's page.
The extra credit assignment is due on 4/21 at 5pm.
This Week's Dialogue Group Meeting
Find at least one hour to meet with your group to discuss the prompt of the week: What does your set of chosen virtues look like in the context of artificial intelligence? This is the conceptual inquiry piece of your design pattern creation process.
Schedule your group meeting for next week. The new prompt is: Identify key features of four different AI algorithms / platforms (your choice) and consider if these features uphold or violate your chosen virtues. This is the first part of the technical inquiry piece of your design pattern creation process.
Once you have completed the readings, fill out the following quiz. If you can't see the form below, try this direct link.