Protected Attributes and 'Fairness through Unawareness,' Exploring Fairness in Machine Learning

TL;DR

Machine learning has the potential to embed bias and unintentionally discriminate, even in large organizations like Amazon and Facebook.

Transcript

[MUSIC PLAYING] MIKE TEODORESCU: Hello, and welcome to this module on protected attributes and fairness through unawareness. My name is Mike Teodorescu. I'm an assistant professor of information systems at Boston College, as well as a visiting scholar at MIT D-Lab. What this module will cover will be examples of laws that codify protected attribute... Read More

Key Insights

• 🎰 Machine learning has both risks and opportunities, as it can automate tasks and reduce costs but can also increase biases.
• 🤕 Protected attributes like race, gender, and age are labeled as sources of social bias and discrimination is illegal in certain contexts.
• 🎰 Despite legal frameworks, machine learning algorithms can still unintentionally embed bias and perpetuate inequality.
• 🤮 Fairness through unawareness, which omits protected attributes, may still include other variables that are correlated with those attributes and lead to biased outcomes.
• 🎰 Data quality, individual biases, and hidden correlations in input data all contribute to the potential for bias in machine learning algorithms.
• 🌥️ Large organizations like Amazon and Facebook have encountered challenges in ensuring fairness in their machine learning systems.
• ❓ Unawareness as a fairness method can result in unintended discrimination and legal consequences.

Questions & Answers

Q: What are some examples of large organizations unintentionally discriminating through machine learning?

Amazon used a machine learning algorithm to screen resumes and discovered that it was discriminating against female applicants. Facebook has also been accused of violating the Fair Housing Act with its machine learning algorithms for displaying ads.

Q: Why is fairness through unawareness not effective in eliminating bias?

Fairness through unawareness ignores protected attributes but may still include other variables that are highly correlated with those attributes. For example, gender was found to cause unintentional bias in Google's advertising system, resulting in targeted ads for high-income jobs being shown more frequently to men.

Q: What are the risks to an organization that chooses unawareness as a fairness method?

By ignoring protected attributes and focusing solely on other variables, the organization may perpetuate hidden correlations and biases in their data. This can lead to discriminatory outcomes and potential legal consequences.

Q: What variables might lead to biased predictions in a machine learning hiring system?

Variables such as college attended, hometown, or other resume indicators that remain unprotected may still be highly correlated with protected attributes like gender and race. This can result in biased predictions and unfair hiring practices.

Summary & Key Takeaways

• Certain individual attributes like race, religion, gender, and socioeconomic status are considered protected attributes and are sources of social bias.

• Despite laws that prohibit discrimination based on protected attributes, machine learning algorithms can still unintentionally embed bias.

• Fairness through unawareness, which involves leaving out protected attributes, may perpetuate inequality by including other attributes that are correlated with the protected attributes.