1 of 13

Intersectionality in AI Fairness

Jennifer Mickel

​

Polymathic Scholar

2 of 13

Agenda

HOW WE APPROACH THIS

01

What is intersectionality?

02

Bias in AI systems

03

How can we make AI more fair?

04

Proposed tool

3 of 13

What is Intersectionality?

  • The idea that people’s identities can interact in unique ways contribute to the oppressions and privileges they experience
  • Identity can change depending on the cultural context

4 of 13

AI Systems Exhibit Bias

Language Models

Show social norm biases such as gender stereotypes

Hiring Algorithms

Discriminate against women

EXAMPLES

Facial Recognition Systems

Showcase bias against women and darker-skinned people

5 of 13

Facial Recognition Precision (PPV)

Lighter Skinned Males

99.6%

Darker Skinned Males

93.8%

Lighter Skinned Females

95.1%

Darker Skinned Females

80.5%

Buolamwini & Gebru, 2018

6 of 13

How Does this Happen?

7 of 13

AI Model Development Pipeline

STAGES

DEVELOPMENT

TRAINING/

Testing

DATA/SET(S)

DEPLOYMENT

8 of 13

DATASET DEVELOPMENT

ANNOTATION

Are the annotations correct?

DISTRIBUTION

What is the distribution of examples in the dataset?

REPRESENTATION

What perspectives are represented in the dataset?

BIAS IN DATASETS

9 of 13

MODEL DEVELOPMENT

AVENUES BIAS CAN ENTER

ARCHITECTURE

Inductive biases

01

TRAINING

Prioritizing accuracy at the expense of fairness

02

TESTING

Neglecting performance metrics

03

10 of 13

What Can We Do About This?

Consider Intersectionality

Consult

Tools

Development

Datasets

11 of 13

HOW CAN WE HELP DO THAT?

OPEN-SOURCE

SEARCHABLE

CONNECTING

  • Users with different identities and experiences can add missing identities and domains to the Intersectionality Tool
  • Users search based on the dataset and/or model domain
  • Results showcase relevant identities
  • Contact information about relevant stakeholders is available
  • Users can connect with these stakeholders

INCREASING INTERSECTIONAL INSIGHTS (I 3) TOOL

12 of 13

RECAP:

PROBLEMS/WHAT CAN WE DO?

INTERSECTIONALITY

CONSIDER

DURING

  1. Dataset development
  2. Model architecture development
  3. Training/testing of the model
  4. Deployment and auditing

BIAS

Can arise from any point of the model development process

HARM

Can arise from AI systems

13 of 13

Thank You!!! Acknowledgements and Sources

I would like to thank Dr. Maria De-Arteaga, Dr. Tina Peterson, Dr. Sina Fazelpour, and Dr. Rebecca Wilcox for their guidance and support throughout the thesis writing process as well as everyone else who has supported me through this process!

​

Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A.,

Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., … Amodei, D. (2020). Language models are few-shot learners. Neural Information Processing Systems, 33, 1-25. https://proceedings.neurips.cc/paper/2020/hash/1457c0d6bfcb4967418bfb8ac142f64a-Abstract.html

Buolamwini, J., & Gebru, T. (2018). Gender shades: Intersectional accuracy disparities in

commercial gender classification. Proceedings of the 1st Conference on Fairness, Accountability and Transparency, 77–91. http://proceedings.mlr.press/v81/buolamwini18a/buolamwini18a.pdf

Cheng, M., De-Arteaga, M., Mackey, L., & Kalai, A. T. (2021). Social norm bias: Residual

harms of fairness-aware algorithms (3rd ed.). arXiv, n.p. https://doi.org/10.48550/arXiv.2108.11056

(see thesis for full list of sources)

This Slides presentation is from Slidesgo.

​