Computing bias

This sprint focuses on the reasons behind bias in computing and how we can address them.

A 2019 study by the National Institute of Standards and Technology (NIST) uncovered implicit biases in facial recognition algorithms.
NIST studied nearly 200 algorithms, highlighting embedded biases in computing innovations.
Computing innovations can reflect and propagate existing human biases due to biased algorithms or the data they are trained on.

Definition: Bias refers to systematic and often unfair errors in computer systems and algorithms resulting from the data used for training or design choices made.
Impact: Can lead to discriminatory outcomes, perpetuating and exacerbating existing inequalities and injustices.

Understanding AI Bias: What is AI bias, and how does it manifest in technology?
Sources of Bias: How do AI systems acquire biases from human data? What role does training data play in AI bias development?
Implications of AI Bias: How might AI bias affect decision-making processes across industries?

Facial Recognition: Algorithms have shown lower accuracy for individuals with darker skin tones, resulting in higher rates of false positives and negatives.
Hiring Algorithms: These may demonstrate bias against women or people of color, consequently reducing job opportunities and perpetuating systemic inequalities.
- Example: In 2018, a corporation developed an AI hiring tool biased against women, which was later abandoned.

Biased Training Data: Algorithms trained on biased datasets lead to biased outcomes.
- Example: A facial recognition algorithm primarily trained on images of white men performed poorly for diverse races and genders (Joy Buolamwini's research).
Biased Algorithms: Systematic errors within algorithm design and training processes perpetuate bias.
Biased Design Assumptions: Incorrect assumptions can lead to biased technology outcomes.
Lack of Diversity in Development Teams: Narrow perspectives in design processes can overlook diverse user needs, exemplifying the "mirror effect."
- Example: Voice recognition tech's struggle with non-native English speakers found that misunderstood rates were nearly double for them compared to native speakers.

Ethical Considerations: Developers and organizations must address bias in AI, ensuring fairness and equity.
Strategies:
- Ensure diverse data collection.
- Audit algorithms regularly.
- Promote transparency and explainability in algorithms.
- Encourage diversity among design and development teams.
- Implement robust testing procedures for bias.

Ensuring Algorithmic Transparency:
- Example: Google's "Why this ad?" feature enhances understanding of algorithmic decisions, reducing perceived biases.
Advocating for Diverse Development Teams: Diverse teams challenge biases effectively.
Regular Algorithm Review and Audit: Scrutiny ensures fairness; for instance, Google regularly audits its search algorithm for bias.
Using Diverse and Representative Data: Broad datasets minimize bias; essential in facial recognition tech.
Implementing Bias Testing: Tools like IBM's AI Fairness 360 test biases in algorithms, paving pathways for fairer outputs.

Challenge: Bias can perpetuate, suggesting extremist content.
Solutions:
- Address user engagement bias by not overly relying on clicks and likes.
- Utilize diverse training data to mitigate bias.
- Implement careful content prioritization based on quality, not just engagement.
- Establish feedback mechanisms for users to report inappropriate content.

Challenge: Potential biases based on socio-economic status, gender, race, age, and location.
Solutions:
- Include diverse training data.
- Regular audits and human oversight during recruitment.
- Ensure transparency about AI decision-making processes.

Challenge: Biased algorithms in identifying gender/skin tone accurately.
Solutions:
- Utilize diverse training datasets.
- Regular audits using benchmark datasets for performance evaluation.
- Implement bias detection methods and involve human experts during development and validation.