Regression to the Wrong Answer

A new candidate has 6 years of experience and graduated from a top-ranked college. Based only on the data above, would you predict they get hired?

• Yes

• No

• Can't tell from the data

Model B has a higher R² than Model A. What does this tell us?

• Model B is better and should be used for hiring decisions

• Model B makes more accurate predictions based on its training data, but accuracy alone doesn't make a model fair or appropriate

• Model B is broken and needs to be fixed

• R² doesn't matter — only fairness matters

Amazon's AI hiring tool gave lower scores to résumés that mentioned women's organizations. What best explains why?

• The AI was programmed by engineers who were biased against women

• The AI learned from historical hiring data in which men were hired more often, so it associated male-coded signals with success

• The AI made a random error that happened to disadvantage women

• Women's organizations are less impressive than men's organizations

A company removes "gender" from their hiring model's input data to make it fairer. A researcher points out this may not fully solve the problem. Why?

•  Because fairness and accuracy cannot both be achieved at the same time

• Because gender is actually important for predicting job performance

• Because other variables in the dataset, like college attended or job title, may still correlate with gender, allowing the model to use gender indirectly

•  Because removing variables always makes models less accurate

Which of the following best describes the main argument of this passage?

• AI models should not be used for hiring decisions under any circumstances

• Bias in AI is always the result of intentional decisions by engineers

• The solution to algorithmic bias is to use more data

• A mathematically accurate model can still produce unfair outcomes if it is trained on biased historical data

In your own words, explain why a model trained on biased data might be described as "laundering" bias through math. What makes this different from, or similar to, human bias?