Embedded Files
Bias in AI
Bias in AI
Garbage In = Garbage Out
Dataset Biases
Dataset Biases
Dataset biases refer to biases caused by using biased or flawed data during the training of an AI model.
Example of Harmful AI Bias
Example of Harmful AI Bias
"In the Image Playground app, you can use Apple Intelligence to combine concepts, text descriptions, and people from your photo library to create images within seconds."
Input Picture
Algorithmic Biases
Algorithmic Biases
Algorithmic biases refer to biases that arise from choices made during the development and implementation of the AI model's algorithm.
Optimization Bias
Optimization Bias
Example:
A healthcare AI trained to reduce costs may decide to save money at the expense of patient care (like denying treatments that are necessary but expensive).
When AI only focuses on one goal, it can miss the bigger picture and lead to unfair outcomes.
Interaction Bias
Interaction Bias
Example:
A user searches, “Are certain people better at math than others?”
They click on a link with a biased claim like: “Men are naturally better at math than women.”
The AI notices that click and interprets it as a “preference." During the user's next search the AI is more likely to show similar results higher in the list, not because they’re more accurate, but because it engaged the user.
The more users interact with biased content, the more the AI may promote it, creating a feedback loop.
AI Bias in the News
AI Bias in the News
University of Washington research found significant racial, gender and intersectional bias in how three state-of-the-art large language models ranked resumes. The system preferred white names 85% of the time versus 9% for black associated names (Milne, 2024).
AI models consistently assigned speakers of African American English to lower-prestige jobs and issued more convictions in hypothetical criminal cases—and more death penalties. This bias is also not exclusive to AAE; researchers conducted preliminary testing of Indian and Appalachian English and discovered similar biases (Hofmann, Kalluri, Jurafsky, & King, 2024).
Ai models are often used in medical diagnosis, particularly in analyzing images like x-rays. These models have been shown to accurately predict a patient's race based on the x-ray alone. Research is showing that the models may be using "demographic shortcuts" to make diagnostic evaluations (Trafton, 2024).
Next: Accuracy Issues
Page updated
Google Sites
Report abuse