Ethics & Bias in Social Media

🎯 Learning Objectives

• Identify sources of bias in social media data and algorithms
• Apply fairness metrics to evaluate algorithmic systems
• Understand IRB requirements for social media research
• Evaluate ethical dilemmas in platform design
• Develop responsible data practices for social media research

1. Algorithmic Bias

1.1 Intuition

1.2 Sources of Bias

1.3 Worked Example: Gender Bias in Word Embeddings

python
Run Inline
import gensim.downloader as api
glove = api.load("glove-twitter-25")
# Test gender bias
pairs = [
    ("doctor", "nurse"),
    ("computer", "cooking"),
    ("engineer", "teacher"),
    ("programmer", "homemaker"),
]
for male_job, female_job in pairs:
    # Measure stereotypical association
    similarity_m = glove.similarity(male_job, "man")
    similarity_f = glove.similarity(male_job, "woman")
    diff = similarity_m - similarity_f
    print(f"{male_job}: man={similarity_m:.3f}, woman={similarity_f:.3f}, bias={diff:.3f}")

pseudo
doctor: man=0.412, woman=0.315, bias=0.097
computer: man=0.352, woman=0.283, bias=0.069
engineer: man=0.401, woman=0.298, bias=0.103
programmer: man=0.387, woman=0.271, bias=0.116

2. Fairness Metrics

2.1 Defining Fairness

2.2 The Impossibility Theorem

python
Run Inline
def demonstrate_fairness_impossibility():
    """
    Show that demographic parity and equal opportunity conflict
    when base rates differ.
    """
    # Group A: 10% hate speech, Group B: 1% hate speech
    groups = {
        'A': {'total': 1000, 'hate_speech': 100},  # 10%
        'B': {'total': 1000, 'hate_speech': 10},     # 1%
    }
    # Model that catches 80% of hate speech in both groups
    tp_rate = 0.8
    fp_rate = 0.05
    for group, data in groups.items():
        tp = int(data['hate_speech'] * tp_rate)
        fp = int((data['total'] - data['hate_speech']) * fp_rate)
        total_flagged = tp + fp
        print(f"Group {group}:")
        print(f"  Total flagged: {total_flagged} ({total_flagged/data['total']*100:.1f}%)")
        print(f"  TP: {tp}, FP: {fp}")

pseudo
Group A: Total flagged: 125 (12.5%), TP: 80, FP: 45
Group B: Total flagged: 57 (5.7%), TP: 8, FP: 49

2.3 Bias Mitigation Strategies

3. Research Ethics

3.1 Historical Ethical Failures

3.2 IRB Framework

3.3 Key Ethical Principles

4. Platform Ethics

4.1 Content Moderation Dilemmas

4.2 The Filter Bubble

1. User clicks on content A
2. Algorithm learns user likes content A
3. Algorithm shows more content similar to A
4. User never sees content B (opposing views)
5. User's beliefs become more extreme

5. Common Pitfalls

Pitfall 1: Assuming "Public" Data Means "No Ethics Required"

Pitfall 2: Confusing Statistical Parity with Fairness

Pitfall 3: Ignoring Feedback Loops

6. Key Concepts Reference

7. 📝 Practice Questions

Q1: A hate speech detection model has 90% accuracy overall but 60% accuracy for African American English (AAE) tweets. What type of bias is this?

Answer: This is representation bias combined with measurement bias. AAE is underrepresented in training data (representation bias), and the features used to detect hate speech may misclassify legitimate AAE expressions as hate speech (measurement bias). The model's high overall accuracy masks this disparity — a classic case where aggregate metrics hide per-group failures. Q2: Why is it mathematically impossible to satisfy both demographic parity and equal opportunity when base rates differ?

Answer: Demographic parity requires equal prediction rates (P(Ŷ=1|A) = P(Ŷ=1|B)). Equal opportunity requires equal true positive rates (P(Ŷ=1|Y=1,A) = P(Ŷ=1|Y=1,B)). If base rates differ (P(Y=1|A) ≠ P(Y=1|B)), satisfying both creates a contradiction: the prediction rate depends on both the true positive rate and the base rate. You can match one but not both simultaneously unless base rates are identical. Q3: You scrape 10,000 public tweets about a sensitive health topic. What ethical obligations do you have?

Answer: (1) Check platform terms of service — does scraping violate them? (2) Anonymize user handles and any identifiable information before analysis. (3) Do not quote individual tweets in a way that identifies users. (4) Consider whether users would expect their tweets to be used for this purpose. (5) Store data securely with access controls. (6) Consult IRB — some institutions consider this human subjects research. (7) Consider potential harm if findings could stigmatize the health condition. Q4: Explain the Facebook Emotional Contagion study and its ethical problems.

Answer: Facebook manipulated the news feeds of 689,003 users to show either more positive or more negative content, measuring whether this affected users' own posting sentiment. Ethical problems: (1) No informed consent — users didn't know they were in an experiment. (2) No IRB approval — Facebook's internal review was insufficient. (3) Potential harm — manipulating users' emotions could cause psychological distress. (4) Deception — users were misled about why they saw certain content. The study was published in PNAS and sparked major debate about research ethics in social media. Q5: What is a filter bubble and how does algorithmic amplification contribute to it?

Answer: A filter bubble is a state of intellectual isolation resulting from algorithmic personalization. The algorithm shows users content they're likely to engage with (based on past behavior). If a user clicks political content from one perspective, the algorithm learns to show more of that perspective and less from opposing views. Over time, the user sees an increasingly narrow range of content, which can polarize their beliefs. The algorithm optimizes for engagement, not information diversity, so it naturally creates filter bubbles. Q6: A recommendation system suggests extreme content because it increases engagement. How would you fix this?

Answer: (1) Change the optimization metric — instead of pure engagement, use a weighted metric that penalizes content beyond a certain extremeness threshold. (2) Add diversity constraints — ensure recommendations include diverse perspectives. (3) Implement content scoring — give lower scores to content flagged as potentially problematic. (4) Audit regularly — check that recommendations don't push users toward extreme content. (5) Give users control — allow them to adjust their recommendation preferences. The key insight is that engagement-optimized systems naturally find extreme content (which is highly engaging) — you must explicitly counter this. Q7: Your model for content recommendation has different accuracy for different demographic groups. List three approaches to mitigate this.

Answer: (1) Pre-processing: Collect more representative training data from underrepresented groups. Augment the minority group data. (2) In-processing: Add fairness constraints to the loss function (e.g., penalize differences in error rates across groups). Use adversarial debiasing where a secondary model tries to predict the protected attribute from the predictions. (3) Post-processing: Calibrate thresholds per group so that error rates are balanced. Adjust predictions using techniques like equalized odds post-processing. Q8: What's the difference between "fairness through unawareness" and "fairness through awareness"?

Answer: "Fairness through unawareness" removes protected attributes (race, gender) from the model input, assuming that if the model doesn't see these attributes, it can't discriminate. This fails because other features (zip code, name, interests) can proxy for protected attributes. "Fairness through awareness" explicitly measures and addresses disparities, using techniques like demographic parity or equal opportunity constraints. Awareness is more effective because it acknowledges that bias exists even without explicit protected attributes.

8. 🔗 Cross-References

• Week 4 - Cyber Crime: Ethical hacking
• Week 9 - Privacy Papers: Research ethics in papers
• Week 11 - Comp Social Science: Ethical computational research
• BSCS4021 (Advanced Algorithms): Algorithmic analysis Join Discord PreviousPrivacy PapersNextComp Social Science