Worksheet: Bias Detection in Text

Background

Bias in text can perpetuate stereotypes and unfair treatment. Detecting bias is crucial for building fair AI systems and understanding media representation.

Textual Bias: Language that unfairly represents, stereotypes, or marginalizes individuals or groups based on attributes like gender, race, age, or socioeconomic status.

Types of Bias

Gender Bias Stereotyping based on gender/sex

Racial/Ethnic Bias Prejudice based on race or ethnicity

Age Bias Discrimination based on age

Socioeconomic Bias Prejudice based on class/wealth

Ability Bias Bias against people with disabilities

Part 1: Identifying Explicit Bias

Classify each example and identify the type of bias present.

"For a woman, she's surprisingly good at math."

Gender bias Racial bias Age bias Socioeconomic bias No bias

"The elderly driver was, predictably, going 20 under the speed limit."

Gender bias Racial bias Age bias Socioeconomic bias No bias

Question 1

What makes a compliment biased (as in the first example)?

Part 2: Subtle and Implicit Bias

"The doctor examined his patient carefully. Meanwhile, the nurse prepared her supplies."

Biased (assumes doctor=male, nurse=female) Neutral (just describes specific individuals) Depends on context Ambiguous - cannot determine

"The articulate young Black man impressed the interviewers."

Biased (implies surprise at articulateness) Neutral (just descriptive) Positive (compliment) Depends on context

Question 2

Why might calling someone "articulate" be biased in certain contexts?

Part 3: Framing and Word Choice

The same event can be described in biased or neutral ways.

Version A: "Protesters clashed with police during the demonstration."

Version B: "Rioters attacked police officers during the violent gathering."

Version C: "Activists were confronted by law enforcement during the peaceful march."

Question 3

Rate the bias in each version:

Version	Bias Direction	Severity
Version A
Version B
Version C

Identify the specific word choices that create bias:

Part 4: Statistical vs. Individual Bias

"Men are, on average, taller than women."

Question 4a

Is this statement biased?

Yes, reinforces gender stereotypes No, it's a statistical fact Depends on context/usage

"We need a tall person for this job, so let's hire a man."

Question 4b

Is this statement biased?

Yes, applies group statistics to individuals No, it's statistically reasonable Depends on the specific job

What's the difference between these two examples?

Part 5: Bias in AI-Generated Text

Prompt: "Write a story about a CEO."

AI Output: "John walked into his corner office, adjusting his tie. As CEO of the company his father had built, he felt the weight of responsibility on his shoulders. His secretary, a young woman named Lisa, brought him his morning coffee..."

Question 5

Identify potential biases in this AI-generated text:

CEO assumed to be male Secretary assumed to be female Age bias (young secretary) Inherited wealth assumption Professional dress stereotypes

Why do AI systems often produce biased outputs?

Part 6: Annotating a Full Passage

"The tech industry has long struggled with diversity. Despite efforts to recruit more women and minorities, the field remains dominated by young white and Asian men. Critics argue that the 'culture fit' hiring practice often favors candidates who look and think like existing employees. Meanwhile, older workers find themselves pushed out as companies chase the next young genius. Some suggest that the industry's obsession with youth ignores the value of experience."

Question 6

Analyze this passage for bias:

What biases does the passage DISCUSS (topic of the text)?

Does the passage itself EXHIBIT any bias (how it's written)?

How should annotation guidelines distinguish between discussing bias and exhibiting bias?

Part 7: Group Discussion

Question 7

Compare your annotations with your group. Where did you disagree?

Whether bias is present Type of bias Severity of bias Subtle vs. explicit Context interpretation

Describe one specific disagreement:

Part 8: Reflection

Question 8

Why is bias detection difficult?

Bias perception is subjective Much bias is subtle/implicit Context is crucial Annotator background matters Social norms evolve All of the above

Key Takeaway

Bias is often in the eye of the beholder, but annotation can still be systematic.

Explicit bias is easier to identify than implicit bias
Word choice and framing reveal underlying assumptions
Discussing bias vs. exhibiting bias are different
Annotator diversity is crucial for comprehensive coverage
Guidelines must acknowledge subjectivity while providing clear criteria