DAT566: Module 8 Ethics and Sustainability of AI

What are the core principles of AI Ethics?

Transparency and Accountability

What was the first widely acknowledged case of AI discrimination?

1982 St. George’s Hospital Medical School

What is technochauvinism?

The belief that technological solutions are always superior, neutral, or more efficient than human or social ones

Which aspects should a sociotechnical solution focus on?

Social, technical, and legal solutions working together

What are considered the "hidden costs" of the AI supply chain?

The precarious, low-paid human labor involved in data annotation.

The e-waste generated from obsolete AI hardware.

The concept of "Data Colonialism" describes a power imbalance in the AI industry. Which statement best summarizes this concept?

The extraction of data and cheap labor from the Global South to build technologies that primarily concentrate wealth and power in the Global North.

What did the ProPublica study on COMPAS bring to light?

An algorithmic bias that disproportionately labeled black individuals as having a high risk of recidivism

How is power redistributed and participation reclaimed?

Activism, critical public literacy, participatory design

What pre-processing strategies are bias mitigation strategies?

Re-weighing and data transformation

What in-processing strategies are bias mitigation strategies?

Fair regularization and constraint based optimization

What post-processing strategies are bias mitigation strategies?

Calibration and threshold adjustment

What are some auditing tools?

Identify disparities across demographic groups, apply fairness metrics, evaluate trade-offs, test mitigation strategies

What scenario best illustrates regulatory capture in the AI industry?

An AI company lobbies for regulation that limits competitors.

What are the two aspects of AI sustainability?

Social (people): ensuring fairness, equity, and respect for human rights; avoiding exploitation

Environmental (planet): Minimizing the carbon footprint, energy consumption, and e-waste

What is the productivity bandwagon?

When technological advances lead to widely shared prosperity by increasing wages and opportunities for the majority of workers

How can the productivity bandwagon come to be?

Strong labor unions, government regulation, investments in education

What analogy did Kate Crawford make to the AI industry?

The mining industry

What are “Green AI” strategies?

Algorithmic efficiency: the most effective strategy is to be smarter with code

1. Model compression: pruning, quantization, knowledge distillation

2. Transfer learning and fine tune pre trained models

3. Re-thing your needs: Do you need largest model available?

BONUS: Which statement best reflects Skiena’s treatment of methods and tools across the data science pipeline, from data acquisition to visualization and modeling?

Advances in algorithms replaces the importance of data quality and exploratory analysis.

Data set bias is overrated.

Effective data science requires adapting methods based on data properties, domain constraints, and evaluation goals.

Standardized workflows guarantee reproducibility and optimal performance across problems.

Visualization primarily serves to communicate results after modeling is complete.

Effective data science requires adapting methods based on data properties, domain constraints, and evaluation goals.

BONUS: A data science team builds a model that achieves strong cross-validation performance and clean visualizations. However, when deployed, the model leads to poor real-world decisions.

According to Skiena’s overarching view of data science, which explanation is most consistent with the failure?

Cross-validation is unreliable for real-world problems.

The model should have used a more sophisticated algorithm.

The evaluation metrics failed to capture the real decision objective and costs.

Visualization biased stakeholders toward incorrect conclusions.

Real-world performance is not worth striving for.

The evaluation metrics failed to capture the real decision objective and costs.