Evaluating and mitigating social bias

Generative AI Concepts

Daniel Tedesco

Data Lead, Google

What do we mean by social bias?

$$ $$

Systematic unfairness in generative AI

Serious societal consequences
Fairness can be subjective
Focus on broadly shared values

$$ $$

A job interview.

Where bias appears

Training data
The model itself
How the model is used

Unbalanced weighing scales

Bias in data

Skewed or unrepresentative information in the training dataset

Two training funnels. One shows purple squares going into training and only purple squares being generated. The other shows a variety of colored shapes going into training and a variety of shapes being generated

Bias in models

Pursuing goals that result in biased outcomes

Photorealistic robot politician giving a speech at the UN

Bias in use

Applying AI in wrong or malicious ways

Laptop showing a video with a giant "FAKE" label on it.

Identifying bias in data and models

Representation analysis compares how the model refers to different groups
Fairness metrics evaluate models for equal treatment, opportunity, and accuracy across groups
Human audits ask real people to review a model's outputs to identify bias

Mitigating bias in data and models

Diversify data collection
Adjust model to prioritize different data
Adversarial training
Continuous improvement

A collection of shapes of different colors and a representative sample of them selected

Let's practice!

Generative AI Concepts